Foundations of Systems Biology

  • 56 252 10
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up

Foundations of Systems Biology

edited by Hiroaki Kitano The MIT Press Cambridge, Massachusetts London, England c 2001 Massachusetts Institute of T

1,344 215 4MB

Pages 290 Page size 336 x 430.56 pts Year 2005

Report DMCA / Copyright


Recommend Papers

File loading please wait...
Citation preview

Foundations of Systems Biology

edited by Hiroaki Kitano

The MIT Press Cambridge, Massachusetts London, England

c 2001 Massachusetts Institute of Technology  All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from the publisher. This book was set in Palatino by the author using the LATEX document preparation system. Printed on recycled paper and bound in the United States of America. Library of Congress Cataloging-in-Publication Data Foundations of systems biology edited by Hiroaki Kitano. - 1st ed. Included bibliographical references (p.). ISBN 0-262-11266-3 (hc.: alk. paper) 1. Biological systems-Research-Methodology 2. Biological systems-Research-Case studies. I. Kitano, Hiroaki, 1961QH313.F662002 573-dc21 2001042807


Contributors Preface 1

Systems Biology: Toward System-level Understanding of Biological Systems Hiroaki Kitano




Automatic Acquisition of Cell Lineage through 4D Microscopy and Analysis of Early C. elegans Embryogenesis Shuichi Onami, Shugo Hamahashi, Masao Nagasaki, Satoru Miyano, and Hiroaki Kitano

ix xiii 1





The DBRF Method for Inferring a Gene Network from Large-Scale Steady-State Gene Expression Data Shuichi Onami, Koji M. Kyoda, Mineo Morohashi, and Hiroaki Kitano





The Analysis of Cancer Associated Gene Expression Matrices Mattias Wahde and Zoltan Szallasi


Automated Reverse Engineering of Metabolic Pathways from Observed Data by Means of Genetic Programming John R. Koza, William Mydlowec, Guido Lanza, Jessen Yu, and Martin A. Keane







Automatic Model Generation for Signal Transduction with Applications to MAP-Kinase Pathways Bruce E. Shapiro, Andre Levchenko, and Eric Mjolsness


Modeling Large Biological Systems From Functional Genomic Data: Parameter Estimation Pedro Mendes





Towards a Virtual Biological Laboratory Jorg ¨ Stelling, Andreas Kremling, Martin Ginkel, Katja Bettenbrock and Ernst Dieter Gilles


Computational Cell Biology — The Stochastic Approach Thomas Simon Shimizu and Dennis Bray


Computer Simulation of the Cell: Human Erythrocyte Model and its Application Yoichi Nakayama and Masaru Tomita






Constructing Mathematical Models of Biological Signal Transduction Pathways: An Analysis of Robustness Tau-Mu Yi




The ERATO Systems Biology Workbench: An Integrated Environment for Multiscale and Multitheoretic Simulations in Systems Biology Michael Hucka, Andrew Finney, Herbert Sauro, Hamid Bolouri, John Doyle, and Hiroaki Kitano



Combination of Biphasic Response Regulation and Positive Feedback as a General Regulatory Mechanism in Homeostasis and Signal Transduction Andre Levchenko, Jehoshua Bruck, and Paul W. Sternberg


Distinct Roles of Rho-kinase Pathway and Myosin Light Chain Kinase Pathway in Phosphorylation of Myosin Light Chain: Kinetic Simulation Study Shinya Kuroda, Nicolas Schweighofer, Mutsuki Amano, Kozo Kaibuchi, and Mitsuo Kawato



Contributors Mutsuki Amano Division of Signal Transduction, Nara Institute of Science and Technology.

and Control and Dynamical Systems, California Institute of Technology. Andrew Finney

Katja Bettenbrock

[email protected]

Max-Planck-Institute for Dynamics of Complex Technical Systems.

JST/ERATO Kitano Systems Biology Project, and Control and Dynamical Systems, California Institute of Technology.

[email protected]

Hamid Bolouri

[email protected]

JST/ERATO Kitano Systems Biology Project, and Control and Dynamical Systems, California Institute of Technology, and Division of Biology, California Institute of Technology, and Science and Technology Research Centre, University of Hertfordshire.

Ernst Dieter Gilles

[email protected]

Max-Planck-Institute for Dynamics of Complex Technical Systems. Martin Ginkel

[email protected]

Max-Planck-Institute for Dynamics of Complex Technical Systems. Shugo Hamahashi

Dennis Bray

[email protected]

Department of Zoology, University of Cambridge.

[email protected]

Kitano Symbiotic Systems Project, ERATO, JST, and Department of Computer Science, Keio University.

Jehoshua Bruck

[email protected]

Division of Engineering and Applied Science, California Institute of Technology. John Doyle

[email protected]

Kitano Systems Biology Project, ERATO, JST,

Michael Hucka

[email protected]

JST/ERATO Kitano Systems Biology Project, and Control and Dynamical Systems,

California Institute of Technology. Kozo Kaibuchi Division of Signal Transduction, Nara Institute of Science and Technology, and Department of Cell Pharmacology, Nagoya University. Mitsuo Kawato Kawato Dynamic Brain Project, ERATO, JST, and Human Information Processing Research Laboratories, ATR. Martin A. Keane

[email protected]

Econometric Inc.

Dynamics of Complex Technical Systems. Shinya Kuroda

[email protected]

Kawato Dynamic Brain Project, ERATO, JST, and Division of Signal Transduction, Nara Institute of Science and Technology, Present address: Center for Neurobiology and Behavior, Columbia University. Koji M. Kyoda

[email protected]

Kitano Symbiotic Systems Project, ERATO, JST, and Department of Fundamental Science and Technology, Keio University.

Hiroaki Kitano

[email protected]

Kitano Symbiotic Systems Project, ERATO, JST, and The Systems Biology Institute, and Control and Dynamical Systems, California Institute of Technology, and Sony Computer Science Laboratories, Inc. John R. Koza

[email protected]

Biomedical Informatics, Department of Medicine, Department of Electrical Engineering, Stanford University. Andreas Kremling

[email protected]

Max-Planck-Institute for



Guido Lanza

[email protected]

Genetic Programming Inc. Andre Levchenko

[email protected]

Division of Engineering and Applied Science, California Institute of Technology. Pedro Mendes [email protected]

Virginia Bioinformatics Institute, Virginia Polytechnic Institute and State University. Satoru Miyano

[email protected]

Kitano Symbiotic Systems Project, ERATO, JST, and Human Genome Center, Institute of Medical Science, University of Tokyo.

Control and Dynamical Systems, California Institute of Technology. Eric Mjolsness

[email protected]

Jet Propulsion Laboratory, California Institute of Technology, and Division of Biology, California Institute of Technology, and Kitano Symbiotic Systems Project, ERATO, JST. Mineo Morohashi

[email protected]

Kitano Symbiotic Systems Project, ERATO, JST, and Department of Fundamental Science and Technology, Keio University.

Herbert Sauro

[email protected]

JST/ERATO Kitano Systems Biology Project, and Control and Dynamical Systems, California Institute of Technology. Nicolas Schweighofer Kawato Dynamic Brain Project, ERATO, JST, Present address: Learning Curve. Bruce Shapiro

[email protected]

Jet Propulsion Laboratory, California Institute of Technology.

William Mydlowec

Thomas Simon Shimizu

Genetic Programming Inc.

Department of Zoology, University of Cambridge.

[email protected]

Masao Nagasaki

[email protected]

Kitano Symbiotic Systems Project, ERATO, JST, and Department of Information Science, Human Genome Center, Institute of Medical Science, University of Tokyo. Yoichi Nakayama

[email protected]

Jorg ¨ Stelling

[email protected]

Max-Planck-Institute for Dynamics of Complex Technical Systems. Paul W. Sternberg

[email protected]

Division of Biology and Howard Hughes Medical Institute, California Institute of Technology.

[email protected]

Institute for Advanced Biosciences, Keio University.

Zoltan Szallasi

Shuichi Onami

Department of Pharmacology, Uniformed Services University of the Health Sciences.

[email protected]

Kitano Symbiotic Systems Project, ERATO, JST, and The Systems Biology Institute, and xi


[email protected]

Masaru Tomita

[email protected]

Institute for Advanced Biosciences,

Keio University. Mattias Wahde

[email protected]

Division of Mechatronics, Chalmers University of Technology. Tau-Mu Yi

[email protected]

Kitano Symbiotic Systems Project, ERATO, JST, and Division of Biology, California Institute of Technology. Jessen Yu

[email protected]

Genetic Programming Inc.




Systems biology aims at understanding biological systems at system level. It is a growing area in biology, due to progress in several fields. The most critical factor has been rapid progress in molecular biology, furthered by technologies for making comprehensive measurements on DNA sequence, gene expression profiles, protein-protein interactions, etc. With the ever-increasing flow of biological data, serious attempts to understand biological systems as systems are now almost feasible. Handling this high-throughput experimental data places major demands on computer science, including database processing, modeling, simulation, and analysis. Dramatic progress in semiconductor technologies has led to highperformance computing facilities that can support system-level analysis. This is not the first attempt at system-level analysis; there have been several efforts in the past, the most notable of which is cybernetics, or biological cybernetics, proposed by Norbert Wiener more than 30 years ago. With the limited understanding of biological processes at the molecular level at that time, most of the work was on phenomenological analysis of physiological processes. There have also been biochemical approaches, such as metabolic control analysis, and although restricted to steady-state flow, it has successfully been used to explore system-level properties of biological metabolism. Systems biology, just like all other emerging scientific disciplines, is built on multiple efforts that share the vision. However, systems biology is distinct from past attempts because for the first time we are able to understand biology at the system level based on molecular-level understanding, and to create a consistent system of knowledge grounded in the molecular level. In addition, it should be noted that systems biology is intended to be biology for system-level studies, not physics, systems science, or informatics, which try to apply certain dogmatic principles to biology. When the field has matured in the next few years, systems biology will be characterized as a field of biology at the system level with extensive use of cutting-edge technologies and highly automated high-throughput precision measurement combined with sophisticated computational tools and analysis. Systems biology clearly includes both experimental and computational or analytical studies. However, systems biology is not a mere combination of molecular biology and computing science to reverse

engineer gene networks. Of course, it is one of the topics included, but system-level understanding requires more than understanding structures. Understanding of (1) system structures, (2) dynamics, (3) control methods, and (4) design methods are major four milestones of systems biology research. One of the most important missions of this book, which I tried hard in my chapter, is to define the scope and provide the vision and perspectives of this new born field. I am very pleased to see that interest is rapidly growing among both experimental biologists and those who with computing and engineering backgrounds are seriously interested in biological systems. Not many people understood what I was trying to describe when I was using the term “systems biology” a few years ago, because it was well before human genome sequence to complete, high-throughput experiments were to be considered as realistic option, and it was a new term nobody used before. But, today more and more people are using the concept and the term. Of course, it is the actual research that matters, but the term is also important because it symbolically represents what we are trying to accomplish. Today, we find more and more researchers are getting involved, as well as numbers of research groups and institutions are being formed focusing on systems biology. Fortunately, I managed to convince the Japanese government to support the initiation of a new international conference. The First International Conference on Systems Biology (ICSB2000) was held in Tokyo from November 14–16, 2000, supported by the Japan Science and Technology Corporation, an agency belonging to the Science and Technology Agency of the Japanese government. It was the first international conference that clearly focused on systems biology work. Since then, various international and national conferences, symposiums, and seminars have started organize systems biology sessions. The second conference will be held at the California Institute of Technology in 2001 with the support of Caltech. In fact, Caltech is one of the first institutions that seriously explored systems biology. I still remember the overwhelming reaction when I gave a talk “Perspectives on Systems Biology” at Caltech in March 1998. This book is the first book on systems biology, and consists of papers representing work in the systems biology field. It is loosely based on papers that were presented at ICSB2000. Of course, many research studies related to systems biology are already underway, and I must state that this book is by no means an exhaustive collection of such works. Also, the experimental aspects of systems biology are under-represented here, because many of the projects aiming at next-generation experiments are at their early stage and so are not ready for publication. Nevertheless, the book covers the central themes of systems biology: comprehensive and automated measurements, reverse engineering of gene and metabolic networks from experimental data, software issues, modeling and simulation, and system-level analysis. xiv


Although it it based on a long history, systems biology is a field in its infancy. This book serves two purposes: first, to inform interested researchers on the current state of the research and challenges before us, and second, to be an archival collection of papers to record the initial stage of the research. It is likely, just as in any fast-growing research area, that the technical contents of the book will quickly become obsolete. However, it is often the case that the vision, concept, and philosophy are still valid and add value. I hope the field will quickly grow and flourish beyond its present boundaries, but that the vision outlined herein is enduring. Finally, this book could not have been completed without the support of many people. Mineo Morohashi has done a beautiful job in sorting out and formatting all papers, communicating with authors as well as with The MIT Press, and many other tasks. While I was preoccupied with establishing the new research institution, The Systems Biology Institute (, he did most of the editorial assistant work for me. Thank you, Mineo. Members of the systems biology group of ERATO Kitano Symbiotic Systems Project have been a great help in soliciting papers and, more important, in formulating the basic concepts and vision behind systems biology. John Doyle, Mel Simon, Hamid Bolouri, and Mark Borisuk have been particularly cooperative and supportive. Mario Tokoro and Toshi Doi provided me with a superb research environment at Sony Computer Science Laboratories, Inc. Bob Prior at The MIT Press supported me in this project from the beginning; it was in the summer of 1997 at the International Joint Conference on Artificial Intelligence (IJCAI-97) in Nagoya that Bob walked up to me with a printout of the web page from a talk I had given on systems biology at the University of Cambridge, and asked me to publish this book. I am deeply indebted to all of you.

Hiroaki Kitano Senior Researcher, Sony Computer Science Laboratories, Inc. Director, ERATO Kitano Symbiotic Systems Project, JST, and Director, The Systems Biology Institute Tokyo, Japan




Systems Biology: Toward System-level Understanding of Biological Systems Hiroaki Kitano

Systems biology is a new field in biology that aims at system-level understanding of biological systems. While molecular biology has led to remarkable progress in our understanding of biological systems, the current focus is mainly on identification of genes and functions of their products which are components of the system. The next major challenge is to understand at the system level biological systems that are composed of components revealed by molecular biology. This is not the first attempt at system-level understanding, since it is a recurrent theme in the scientific community. Nevertheless, it is the first time that we may be able to understand biological systems grounded in the molecular level as a consistent framework of knowledge. Now is a golden opportunity to uncover the essential principles of biological systems and applications backed up by indepth understanding of system behaviors. In order to grasp this opportunity, it is essential to establish methodologies and techniques to enable us to understand biological systems in their entirety by investigating: (1) the structure of the systems, such as genes, metabolism, and signal transduction networks and physical structures, (2) the dynamics of such systems, (3) methods to control systems, and (4) methods to design and modify systems for desired properties. This chapter gives an overview of the field of systems biology that will provide a system-level understanding of life. INTRODUCTION

The ultimate goal of biology is to understand every detail and principle of biological systems. Almost fifty years ago, Watson and Crick identified the structure of DNA (Watson and Crick, 1953), thus revolutionizing the way biology is pursued. The beauty of their work was that they grounded biological phenomena on a molecular basis. This made it possible to describe every aspect of biology, such as heredity, development, disease, and evolution, on a solid theoretical ground. Biology became part of a consistent framework of knowledge based on fundamental laws of physics. Since then, the field of molecular biology has emerged and enormous progress has been made. Molecular biology enables us to understand biological systems as molecular machines. Today, we have in-depth un-

derstanding of elementary processes behind heredity, evolution, development, and disease. Such mechanisms include replication, transcription, translation, and so forth. Large numbers of genes and the functions of their transcriptional products have been identified, with the symbolic accomplishment of the complete sequencing of DNA. DNA sequences have been fully identified for various organisms such as mycoplasma, Escherichia coli (E. coli), Caenorhabditis elegans (C. elegans), Drosophila melanogaster, and Homo sapiens. Methods to obtain extensive gene expression profiles are now available that provide comprehensive measurement at the mRNA level. Measurement of protein level and their interactions is also making progress (Ito et al., 2000; Schwikowski et al., 2000). In parallel with such efforts, various methods have been invented to disrupt the transcription of genes, such as loss-of-function knockout of specific genes and RNA interference (RNAi) that is particularly effective for C. elegans and is now being applied for other species. There is no doubt that our understanding of the molecular-level mechanisms of biological systems will accelerate. Nevertheless, such knowledge does not provide us with an understanding of biological systems as systems. Genes and proteins are components of the system. While an understanding of what constitutes the system is necessary for understanding the system, it is not sufficient. Systems biology is a new field of biology that aims to develop a system-level understanding of biological systems (Kitano, 2000). Systemlevel understanding requires a set of principles and methodologies that links the behaviors of molecules to system characteristics and functions. Ultimately, cells, organisms, and human beings will be described and understood at the system level grounded on a consistent framework of knowledge that is underpinned by the basic principles of physics. It is not the first time that system-level understanding of biological system has been pursued; it is a recurrent theme in the scientific community. Norbert Wiener was one of the early proponents of system-level understanding that led to the birth of cybernetics, or biological cybernetics (Wiener, 1948). Ludwig von Bertalanffy proposed general system theory (von Bertalanffy, 1968) in 1968 in an attempt to establish a general theory of the system, but the theory was too abstract to be well grounded. A precursor to such work can be found in the work of Cannon, who proposed the concept of “homeostasis” (Cannon, 1933). With the limited availability of knowledge from molecular biology, most such attempts have focused on the description and analysis of biological systems at the physiological level. The unique feature of systems biology that distinguishes it from past attempts is that there are opportunities to ground system-level understanding directly on the molecular level such as genes and proteins, whereas past attempts have not been able to sufficiently connect systemlevel description to molecular-level knowledge. Thus, although it is not 2

Hiroaki Kitano

the first time that system-level understanding has been pursued, it is the first time to have an opportunity to understand biological systems within the consistent framework of knowledge built up from the molecular level to the system level. The scope of systems biology is potentially very broad and different sets of techniques may be deployed for each research target. It requires collective efforts from multiple research areas, such as molecular biology, high-precision measurement, computer science, control theory, and other scientific and engineering fields. Research needs to be carried out in four key areas: (1) genomics and other molecular biology research, (2) computational studies, such as simulation, bioinformatics, and software tools, (3) analysis of dynamics of the system, and (4) technologies for highprecision, comprehensive measurements. This constitutes a major multi-disciplinary research effort that will enable us to understand biological systems as systems. But what does this mean? “System” is an abstract concept in itself. It is basically an assembly of components in a particular formation, yet it is more than a mere collection of components. To understand the system, it is essential that it can be not only to describe in detail, but also it to comprehend what happens when certain stimuli or disruptions occur. Ultimately, we should be able to design the system to meet specific functional properties. It takes more than a simple in-depth description; it requires more active synthesis to ensure that we have fully understood it. To be more specific, in order to understand biological systems as systems, we must accomplish the following. System Structure Identification: First of all, the structures of the system need to be identified, primarily such as regulatory relationships of genes and interactions of proteins that provide signal transduction and metabolism pathways, as well as the physical structures of organisms, cells, organella, chromatin, and other components. Both the topological relationship of the network of components as well as parameters for each relation need to be identified. The use of highthroughput DNA microarray, protein chips, RT-PCR, and other methods to monitor biological processes in bulk is critical. Nevertheless, methods to identify genes and metabolism networks from these data have yet to be established. Identification of gene regulatory networks1 for multicellular organisms is even more complex as it involves extensive cell-cell communication and physical configuration in three-dimensional space. Structure identification for multicellular organisms inevitably involves not only identifying the structure of gene regulatory networks and metabolism networks, but also understanding the physical structures of whole animals precisely at the 1 In this article, the term “gene regulatory networks” is used to represent networks of gene regulations, metabolic pathways, and signal transduction cascades. 3

Systems Biology: Toward System-level Understanding of Biological Systems

cellular level. Obviously, new instrumentation systems need to be developed to collect necessary data. System Behavior Analysis: Once a system structure is identified to a certain degree, its behavior needs to be understood. Various analysis methods can be used. For example, one may wish to know the sensitivity of certain behaviors against external perturbations, and how quickly the system returns to its normal state after the stimuli. Such an analysis not only reveals system-level characteristics, but also provides important insights for medical treatments by discovering cell response to certain chemicals so that the effects can be maximized while lowering possible side effects. System Control: In order to apply the insights obtained by system structure and behavior understanding, research into establishing a method to control the state of biological systems is needed. How can we transform cells that are malfunctioning into healthy cells? How can we control cancer cells to turn them into normal cells or cause apoptosis? Can we control the differentiation status of a specific cell into a stem cell, and control it to differentiate into the desired cell type? Technologies to accomplish such control would enormously benefit human health. System Design: Ultimately, we would like to establish technologies that allow us to design biological systems with the aim of providing cures for diseases. One futuristic example would be to actually design and grow organs from the patient’s own tissue. Such an organ cloning technique would be enormously useful for the treatment of diseases that require organ transplants. There may be some engineering applications by using biological materials for robotics or computation. By using materials that have self-repair and self-sustaining capability, industrial systems will be revolutionized. This chapter discusses scientific and engineering issues to accomplish in-depth understanding of the system. MEASUREMENT TECHNOLOGIES AND EXPERIMENTAL METHODS

Toward Comprehensive Measurements A comprehensive data set needs to be produced to grasp an entire picture of the organism of interest. For example, the entire sequence has been deduced for yeast, and a microarray that can measure the expression level of all known genes is readily available. In addition, extensive studies of protein-protein interactions using the two-hybrid method are being carried out (Ito et al., 2000; Schwikowski et al., 2000). Efforts to obtain highresolution spatiotemporal localization data for protein are underway. C. elegans is an example of an intensively measured multi-cellular organism. A complete cell lineage has already been identified (Sulston et al., 1983; Sulston and Horvitz, 1977), the topology of the neural system 4

Hiroaki Kitano

has been fully described (White et al., 1986), the DNA sequence has been fully identified (The C. elegans Sequencing Consortium, 1998), a project for full description of gene expression patterns during development using whole-mount in situ hybridization (Tabara et al., 1996) is underway, and the construction of a systematic and exhaustive library of mutants has begun. In addition, a series of new projects has started for measuring neural activity in vivo, and for automatic construction of cell lineage in real time using advanced image processing combined with special microscopy (Yasuda et al., 1999; Onami et al., 2001a). While yeast and C. elegans are examples of comprehensive and exhaustive understanding of biological systems, similar efforts are now being planned for a range of biological systems. Although these studies are currently limited to understanding the components of the system and their local relationship with other components, the combination of such exhaustive experimental work and computational and theoretical research would provide a viable foundation for systems biology. Measurement for Systems Biology Although efforts to systematically obtain comprehensive and accurate data sets are underway, systems biology is much more demanding for experimental biologists than the current practice of biology. It requires a comprehensive body of data and control of the quality of data produced so that it can be used as a reference point of simulation, modeling, and system identification. Eventually, many of the current experimental procedures must be automated to enable high-throughput experiments to be carried out with precise control of quality. Needless to say, not all biological experiments will be carried out in such an automated fashion, for important contributions will be made by small-scale experiments. Nevertheless, large-scale experiments will lay the foundation for system-level understanding. High-throughput, comprehensive, and accurate measurement is the most essential part of biological science. While expectations are high for a computational approach to overcome limitations in the traditional approach in biology, it will never generate serious results without experimental data upon which computational studies can be grounded. For the computational and systems approach to be successful, measurement has to be (1) comprehensive, (2) quantitatively accurate, and (3) systematic. While the requirement for quantitative accuracy is obvious, the other two criteria need further clarification. Comprehensiveness can be further classified into three types: Factor comprehensiveness: Comprehensiveness in terms of target factors that are being measured, such as numbers of genes and proteins. It is important that measurement is carried out intensively for the factors (genes 5

Systems Biology: Toward System-level Understanding of Biological Systems

and proteins) that are related to the central genes and proteins of interest. Unless all genes and proteins are measured, how effectively measurement covers the factors of interest is more important, rather than the sheer number of factors measured. Time-series comprehensiveness: In modeling and analysis of a dynamical system, it is important to capture its behavior with fine-grain time series. Traditional biological experiments tend to measure only the change before and after a certain event. For computational analysis, data measured at a constant time interval are essential in addition to traditional sampling points. Item comprehensiveness: There are cases where several features, such as transcription level, protein interaction, phosphorylation, localization, and other features, have to be measured intensively for the specific target. “Systematic” means that measurement is performed in such a way that obtained data can be consistently integrated. The ideal systematic measurement is simultaneous measurement of multiple features for a single sample. It is not sufficient to develop a sophisticated model and perform analysis using only the mRNA or protein level. Multiple data need to be integrated. Then, each data point has to be obtained using samples that are consistent across various measurements. If samples are prepared in substantially different ways, two data points cannot be integrated. Although this requirement sounds obvious, very few data sets meet these criteria today. These criteria are elucidated in the scenario below with some examples of requirements for experimental data. For example, to infer genetic regulatory networks from an expression profile, comprehensive measurement of the gene expression profile needs to be carried out. Expression data in which only the wild-type is measured is generally unusable for this purpose. The data should have a comprehensive set of deletion mutant and overexpression of each gene. Desirable data sets knock out all genes that are measured in the microarray. If only a limited number of genes can be knocked out due to cost and time constraints, it is critical that genes that are expected to be tightly coupled are intensively knocked out rather than knocking out genes sparsely over the whole possible regulatory network. This is due to computational characteristics of the reverse engineering algorithm that constructs the gene regulatory network from profile data. With such algorithms, sparse data points leave almost unlimited ambiguities on possible network structures. Even with the same number of data points, the algorithm produces much more reliable network hypotheses if measured genes are closely related. This is what is meant by factor comprehensiveness. Time-series comprehensiveness is required for phenomena that are time aligned. Time-series profile data need to be prepared with particular caution in terms of time synchronization of samples to be measured. 6

Hiroaki Kitano

It is often the case in traditional experiments that only two measurement points are set: one before the event and one after the event. For example, many studies in cellular aging research measured the expression level of aging-related genes for young cells, aged cells, and immortalized cells, without measuring changes of expression level on fine-grain time series. In some cases, time-series changes of expression level can be important information to create candidate hypotheses or eliminate possible mechanisms. In addition to measurements before and after a biologically interesting event, measurement should be carried at a constant time interval. Expression profile data that has reliable sample time synchrony and constant time interval is most useful to enable the computational algorithm to reliably fit models and parameters to experimental data. Additional information from protein-protein interactions, such as from yeast two-hybrid experiments, is very useful to infer protein-level interactions that fill the gap between regulation of genes. Both protein interactions and expression profiles should be measured on samples that are prepared identically. This systematic measurement requirement is rather hard to meet currently, because not many research groups are proficient in multiple measurement techniques. After obtaining gene regulatory networks, one needs to find out specific parameters used in the network. To understand dynamics, it is essential that each parameter regarding the network is obtained, so that various numerical simulations and analyses can be performed. Such parameters are binding constant, transcription rate, translation rate, chemical reaction rate, degradation rate, diffusion rate, speed of active transport, etc. Except for special cases, such as red blood cells, these constants are not readily available. Measurement using extracts provides certain information, but often these rate constants vary drastically in vivo. Ideally, comprehensive measurement of major parameters would be performed in vivo, but any measurement that gives reasonable estimates would be of great help. In addition to parameter measurement, it is critically important to measure the phosphorylation state at high resolution. While accuracy is important, the level of accuracy required may vary depending on which part of the system is to be measured. In some parts of the network, the system behavior is sensitive to specific parameter values, and thus has to be measured with high accuracy. In other parts of the system, the system may be robust against fluctuations of large magnitude. In such a case, it may often suffice to confirm that the parameter values fall within the range of stability, instead of obtaining highly accurate figures. The point is that not all parts of the system need to be tuned with the same precision. For example, components for jet engines may have to be produced with high precision, but seat belts do not have to achieve the same precision as jet engine components. In future, the type and accuracy requirements for experiments may be determined by theoretical requirements. 7

Systems Biology: Toward System-level Understanding of Biological Systems

The examples given so far have focused on the process of identification of network structure and parameters that enable simulation and analysis of biochemical networks under the simplified assumption that all materials are distributed homogeneously in the environment. Unfortunately, this is not the case in biological systems. There are subcellular structures and localization of transcription products that cause major diversion from a naive model. Multi-cellular systems require measurement of cell-cell contact, diffusion, cell lineage, gene expression during development, etc. For accurate simulation and analysis, these features have to be measured in a comprehensive, accurate, and systematic manner. We have not developed devices to obtain high-throughput measurements for any of these features. This is a serious issue that has to be addressed. Next-generation Experimental Systems To cope with increasing demands for comprehensive and accurate measurement, a set of new technologies and instruments needs to be developed that offers a higher level of automation and high-precision measurement. First, dramatic progress in the level of automation of experimental procedures for routine experiments is required in order to keep up with increasing demands for modeling and system-level analysis. Highthroughput experiments may turn into a labor-intensive nightmare unless the level of automation is drastically improved. Further automation of experimental procedures would greatly benefit the reliability of experiments, throughput, and total cost of the whole operation in the long run. Second, cutting-edge technologies such as micro-fluid systems, nanotechnology and femto-chemistry may need to be introduced to design and build next-generation experimental devices. The use of such technologies will enable us to measure and observe the activities of genes and proteins in a way that is not possible today. It may also drastically improve the speed and accuracy of measurement for existing devices. In those fields where there are obvious needs, such as sequencing and proteomics, the above goals are already pursued. Beyond the development of high-throughput sequencers using high-density capillary array electrophoresis, efforts are being made to develop integrated microfabricated devices that enable PCR and capillary electrophoresis in a single micro device (Lagally et al., 1999; Simpson et al., 1998). Such devices not only enable miniaturization and precision measurements, but will also significantly increase the level of automation. In the developmental biology of C. elegans, identification of cell lineage is one of the major issues that needs to be accomplished to assist analysis of the gene regulatory network for differentiation. The first attempt to identify cell lineage was carried out entirely manually (Sulston et al., 1983; Sulston and Horvitz, 1977), and it took several years to iden8

Hiroaki Kitano

tify the lineage of the wild type. Four-dimensional microscopy allowed us to collect multi-layer confocal images at a constant time interval, but lineage identification is not automatic. With the availability of exhaustive RNAi knockout for C. elegans, high-throughput cell lineage identification is essential to explore the utility of the exhaustive RNAi. Efforts are underway to fully automate cell lineage identification, as well as threedimensional nuclei position data acquisition (Onami et al., 2001a), fully utilizing advanced image processing algorithms and massively parallel supercomputers. Such devices meet some of the criteria presented earlier, and provide comprehensive measurement of cell positions with high accuracy. With automation, high-throughput data acquisition can be expected. If the project succeeds, it can be used to automatically identify the cell lineage of all RNAi knockout for early embryogenesis. The technology may be augmented, but with major efforts, to automatically detect cell-cell contact, protein localization, etc. Combined with whole mount in situ hybridization and possible future single-cell expression profiling, complete identification of the gene regulatory network for C. elegans may be possible in the near future. SYSTEM STRUCTURE IDENTIFICATION

There are various system structures that need to be identified, such as the structural relationship among cells in the developmental process, detailed cell-cell contact configuration, membrane, intra-cellular structures, and gene regulatory networks. While each of these has significance in corresponding research in systems biology, this section focuses on how the structure of gene regulatory networks can be identified, primarily because it is a subject of growing interest due to the rapid uncovering of genomic information, and it is the control center of various cellular phenomena. In order to understand a biological system, we must first identify the structure of the system. For example, to identify a gene regulatory network, one must identify all components of the network, the function of each component, interactions, and all associated parameters. All possible experimental data must be used to accomplish this non-trivial task. At the same time, inference results from existing experiments should enable the prediction of unknown genes and interactions, which can then be experimentally verified. The difficulty is that such a network cannot be automatically inferred from experimental data based on some principles or universal rules, because biological systems evolve through stochastic processes and are not necessarily optimal. Also, there are multiple networks and parameter values that behave quite similar to the target network. One must identify the true network out of multiple candidates. This process can be divided into two major tasks: (1) network structure identification, and (2) parameter identification. 9

Systems Biology: Toward System-level Understanding of Biological Systems

Network Structure Identification Several attempts have already been made to identify gene regulatory networks from experimental data. They can be classified into two approaches. B OTTOM- UP A PPROACH The bottom-up approach tries to construct a gene regulatory network based on the compilation of independent experimental data, mostly through literature searches and some specific experiments to obtain data of very specific aspects of the network of interest. Some of the early attempts of this approach are seen in the lambda phage decision circuit (McAdams and Shapiro, 1995), early embryogenesis of Drosophila (Reinitz et al., 1995; Hamahashi and Kitano, 1998; Kitano et al., 1997), leg formation (Kyoda and Kitano, 1999a), wing formation (Kyoda and Kitano, 1999b), eye formation on ommatidia clusters and R-cell differentiation (Morohashi and Kitano, 1998), and a reaction-diffusion based eye formation model (Ueda and Kitano, 1998). This approach is suitable when most of the genes and their regulatory relationship are relatively well understood. This approach is particularly suitable for the end-game scenario where most of the pieces are known and one is trying to find the last few pieces. In some cases, biochemical constants can be measured so that very precise simulation can be performed. When most parameters are available, the main purpose of the research is to build a precise simulation model which can be used to analyze the dynamic properties of the system by changing the parameters that cannot be done in the actual system, and to confirm that available knowledge generates simulation results that are consistent with available experimental data. There are efforts to create databases that describe gene and metabolic pathways from the literature. KEGG (Kanehisa and Goto, 2000) and EcoCyc (Karp et al., 1999) are typical examples. Such databases are enormously useful for modeling and simulation, but they must be accurate and represented in such a way that simulation and analysis can be done smoothly. There have been some preliminary attempts to predict unknown genes and their interactions (Morohashi and Kitano, 1998; Kyoda and Kitano, 1999a,b). These attempts manually searched possible unknown interactions to obtain simulation results consistent with experimental data, and did not perform exhaustive searches of all possible spaces of network structures. T OP - DOWN A PPROACH The top-down approach tries to make use of high-throughput data using DNA microarray and other new measurement technologies. Already, 10

Hiroaki Kitano

there have been some attempts to infer groups of genes that have a tight relationship based on DNA microarray data using clustering techniques for the yeast cell cycle (Brown and Botstein, 1999; DeRisi et al., 1997; Spellman et al., 1998) and development of mouse central neural systems (D’haeseleer et al., 1999). Clustering methods are suitable for handling large-scale profile data, but do not directly deduce the network structures. Such methods only provide clusters of genes that are co-expressed in similar temporal patterns. Often, easy-to-understand visualization is required (Michaels et al., 1998). Some heuristics must be imposed if we are to infer networks from such methods. Alternative methods are now being developed to directly infer network structures from expression profiles (Morohashi and Kitano, 1999; Liang et al., 1999) and extensive gene disruption data (Akutsu et al., 1999; Ideker et al., 2000). Most of the methods developed in the past translate expression data into binary values, so that the computing cost can be reduced. However, such methods seriously suffer from information loss in the binary translation process, and cannot obtain the accurate network structure. A method that can directly handle continuous-value expression data was proposed (Kyoda et al., 2000b; Onami et al., 2001b) and reported accurate performance without a serious increase in computational costs. An extension of this method seems to be very promising for any serious research on inference of gene regulatory networks. Genetic programming has been applied to automatically reconstruct pathways and parameters that fit experimental data (Koza et al., 2001). The approach requires extensive computing power, and an example of such is the 1,000 CPU cluster Beowulf-class supercomputer, but the approach has the potential to be practical given the expected speed up of processor chips. Such extensions include the development of a hybrid method that combines the bottom-up and the top-down approach. It is unlikely that no knowledge is available before applying any inference methods; in practical cases, it can be assumed that various genes and their interactions are partially understood, and that it is necessary to identify the rest of the network. By using knowledge that is sufficiently accurate, the possible space of network structures is significantly reduced. One major problem is that such methods cannot directly infer possible modifications and translational control. Future research needs to address integration of the data of the expression profile, protein-protein interactions, and other experimental data. Parameter Identification It is important to identify only the structure of the network, but a set of parameters, because all computational results have to be matched and tested against actual experimental results. In addition, the identified net11

Systems Biology: Toward System-level Understanding of Biological Systems

work will be used for simulating a quantitative analysis of the system’s response and behavioral profile. In most cases, the parameter set has to be estimated based on experimental data. Various parameter optimization methods, such as genetic algorithms and simulated annealing, are used to find a set of parameters that can generate simulation results consistent with experimental data (Hamahashi and Kitano, 1999). In finding a parameter set, it must be noted that there may be multiple parameter sets which generate simulation results equally fitted to experimental data. An important feature of parameter optimization algorithms used for this purpose is the capability to find as many local minima (including a global minima) as possible, rather than finding single global minima. This needs to be combined with a method to indicate specific experiments to identify which one of such parameter sets is the correct parameter set. There are several methods to find optimal parameter sets such as brute force exhaustive search, genetic algorithms, simulated annealing, etc. Most of them are computationally expensive, and have not been considered viable options in the past. But the situation has changed, and it will change in future, too. Although it is important to accurately measure and estimate the genuine parameter values, in some cases parameters are not that critical. For example, it was shown through an extensive simulation that the segment polarity network in Drosophila exhibits a high level of robustness against parameter change (von Dassow et al., 2000). For certain networks that are essential for survival the networks need to be built robust against various changes in parameters to cope with genetic variations and external disturbances. For this kind of network, the essence is embedded into the structure of the network, rather than specific parameters of the network. This is particularly the case when feedback control is used to obtain robustness of the circuits, as seen in bacterial chemotaxis (Yi et al., 2000). Thus, parameter estimation and measurement may need to be combined with theoretical analysis on sensitivity of certain parameters to maintain functionalities of the circuit. SYSTEM BEHAVIOR ANALYSIS

Once we understand the structures of the system, research will focus on dynamic behaviors of the system. How does it adapt to changes in the environment, such as nutrition, and various stimuli? How does it maintain robustness against various potential damage to the system, such as DNA damage and mutation? How do specific circuits exhibit functions observed? To attain system-level understanding, it is essential to understand the mechanisms behind (1) the robustness and stability of the system, and (2) functionalities of the circuits. It is not a trivial task to understand the behaviors of complex biolog12

Hiroaki Kitano

ical networks. Computer simulation and a set of theoretical analyses are essential to provide in-depth understanding on the mechanisms behind the circuits. Simulation Simulation of the behavior of gene and metabolism networks plays an important role in systems biology research, and there are several ongoing efforts on simulator development (Mendes and Kell, 1998; Tomita et al., 1999; Kyoda et al., 2000a; Nagasaki et al., 1999). Due to the complexity of the network behavior and large number of components involved, it is almost impossible to intuitively understand the behaviors of such networks. In addition, accurate simulation models are prerequisite for analyzing the dynamics of the system by changing the parameters and structure of the gene and metabolism networks. Although such analysis is necessary for understanding the dynamics, these operations are not possible with actual biological systems. Simulation is an essential tool not only for understanding the behavior, but also for the design process. In the design of complex engineering systems, various forms of simulation are used. It is unthinkable today that any serious engineering systems could be designed and built without simulation. VLSI design requires major design simulation, thus creating one of the major markets for supercomputers. Commercial aviation is another example. The Boeing 777 was designed based almost entirely on simulation and digital prefabrication. Once we enter that stage of designing and actively controlling biological systems, simulation will be the core of the design process. For simulation to be a viable methodology for the study of biological systems, highly functional, accurate, and user-friendly simulator systems need to be developed. Simulators and associated software systems often require extensive computing power such that the system must run on highly parallel cluster machines, such as the Beowulf PC cluster (Okuno et al., 1999). Although there are some simulators, there is no system that sufficiently covers the needs of a broad range of biology research. Such simulators must be able to simulate gene expression, metabolism, and signal transduction for a single and multiple cells. It must be able to simulate both high concentration of proteins that can be described by differential equations, and low concentration of proteins that need to be handled by stochastic process simulation. Some efforts on simulating a stochastic process (McAdams and Arkin, 1998) and integrating it with high concentration level simulation are underway. In some cases, the model requires not only gene regulatory networks and metabolic networks, but also high-level structures of chromosomes, such as heterochromatin structures. In the model of aging, some attempts are being made to model heterochromatin dynamics (Kitano and Imai, 1998; Imai and Kitano, 1998). Nevertheless, how to model such dynamics 13

Systems Biology: Toward System-level Understanding of Biological Systems

and how to estimate the structure from sparse data and the current level of understanding are major challenges. The simulator needs to be coupled with parameter optimization tools, a hypothesis generator, and a group of analysis tools. Nevertheless, algorithms behind these software systems need to be designed precisely for biological research. One example that has already been mentioned is that the parameter optimizer needs to find as many local minima (including global minima) as possible, because there are multiple possible solutions of which only one is actually used. The assumption that the most optimal solution is used in an actual system does not hold true in biological systems. Most parameter optimization methods are designed to find the global optima for engineering design and problem solving. While existing algorithms provide a solid starting point, they must be modified to suit biological research. Similar arguments apply to other software tools, too. A set of software systems needs to be developed and integrated to assist systems biology research. Such software includes: • a database for storing experimental data, • a cell and tissue simulator, • parameter optimization software, • bifurcation and systems analysis software, • hypotheses generator and experiment planning advisor software, and • data visualization software. How these modules are related and used in an actual work flow is illustrated in Figure 1.1. While many independent efforts are being made on some of this software, so far only limited efforts have been made to create a common platform that integrates these modules. Recently, a group of researchers initiated a study to define a software platform for systems biology. Although various issues need to be addressed for such a software platform, the rest of this section describes some illustrative issues. Efforts are being made to provide a common and versatile software platform for systems biology research. The Systems Biology Workbench project aims to provide a common middleware so that plug-in modules can be added to form a uniform software environment. Beside the software module itself, the exchange of data and the interface between software modules is a critical issue in data-driven research tools. Systems Biology Mark-up Language (SBML) is a versatile and common open standard that enables the exchange of data and modeling information among a wide variety of software systems (Hucka et al., 2000, 2001). It is an extension of XML, and is expected to become the industrial and academic standard of the data and model exchange format. Ultimately, a group of software tools needs to be used for disease modeling and simulation of organ growth and control; this requires a comprehensive and highly integrated simulation and analysis environment. 14

Hiroaki Kitano

Genome/Proteome Database

System Structure Database

Experimental Data Database


Parameter Optimization Module

Experimental Data Interface

System Analysis Module

Visualization Module

Measurement Systems

System Profile Database

(Bifurcation analysis, Flux Balance Analysis, etc.)

Hypotheses Generation Experiment Planning Module

(A) Relationship among Software Tools

Expression profile data Two-hybrid data, RT-PCR data, etc.

Parameter optimizer


Gene regulation network Metabolic cascade network Signal transduction network

Dynamic systems analysis Robustness, stability, bifurcation, etc Design pattern analysis Design pattern extraction

Hypotheses generator

A set of plausible hypotheses Predictions of genes and interactions

Experiment design assistance system

Biological experiments

Experiment plans

(B) Workflow and software tools Figure 1.1 Software tools for systems biology and their workflow

Analysis Methods There have been several attempts to understand the dynamic properties of systems using bifurcation analysis, metabolic control analysis, and sensitivity analysis. For example, bifurcation analysis has been used to understand the Xenopus cell cycle (Borisuk and Tyson, 1998). The analysis creates a phase portrait based on a set of equations describing the essential process of the Xenopus cell cycle. A phase portrait illustrates in which operation point the system is acting, and how it changes behavior if some of the system parameters are varied. By looking at the landscape of the 15

Systems Biology: Toward System-level Understanding of Biological Systems

phase portrait, a crude analysis of the robustness of the system can be made. A group of analysis methods such as flux balance analysis (FBA) (Varma and Palsson, 1994; Edward and Palsson, 1999) and metabolic control analysis (MCA) (Kacser and Burns, 1973; Heinrich and Rapoport, 1974; Fell, 1996) provides a useful method to understand system-level behaviors of metabolic circuits under various environments and internal disruptions. It has been demonstrated that such an analysis method can provide knowledge on the capabilities of metabolic pathways that are consistent with experimental data (Edward et al., 2001). While such methods are currently aiming at analysis of the steady-state behaviors with linear approximation, extention to dynamic and nonlinear analysis would certainly provide a powerful tool for system-level analysis of metabolic circuits. Several other analysis methods have already been developed for complex engineering systems, particularly in the area of control dynamic systems. One of the major challenges is to describe biological systems in the language of control theory, so that we can abstract essential parts of the system within the common language of biology and engineering. ROBUSTNESS OF BIOLOGICAL SYSTEMS

Robustness is one of the essential features of biological systems. Understanding the mechanism behind robustness is particularly important because it provides in-depth understanding on how the system maintains its functional properties against various disturbances. Specifically, we should be able to understand how organisms respond to (1) changes in environment (deprived nutrition level, chemical attractant, exposure to various chemical agents that bind to receptors, temperature) and (2) internal failures (DNA damage, genetic malfunctions in metabolic pathways). Obviously, it is critically important to understand the intrinsic functions of the system, if we are eventually to find cures for diseases. Lessons from Complex Engineering Systems There are interesting analogies between biological systems and engineering systems. Both systems are designed incrementally through some sort of evolutionary processes, and are generally suboptimal for the given task. They also exhibit increased complexity to attain a higher level of robustness and stability. Consider an airplane as an example. If the atmospheric air flow is stable and the airplane does not need to change course, altitude, or weight balance, and does not need to take off and land, the airplane can be built using only a handful of components. The first airplane built by the Wright brothers consisted of only a hundred or so components. The modern jet, such as the Boeing 747, consists of millions of components. One of the 16

Hiroaki Kitano

major reasons for the increased complexity is to improve stability and robustness. Is this also the case in biological systems? Mycoplasma is the smallest self-sustaining organism and has only about 400 genes. It can only live under specific conditions, and is very vulnerable to environmental fluctuations. E. coli, on the other hand, has over 4,000 genes and can live under varying environments. As E. coli evolved it acquired genetic and biochemical circuits for various stress responses and basic behavioral strategies such as chemotaxis (Alon et al., 1999; Barkai and Leibler, 1997). These response circuits form a class of negative feedback loop. Similar mechanisms exist even in eukaryotic cells2 . A crude speculation is that further increases in complexity in multicellular systems toward homo sapiens may add functionalities that can cope with various situations in their respective ecological niche. In engineering systems, robustness and stability are achieved by the use of (1) system control, (2) redundancy, (3) modular design, and (4) structural stability. The hypothesis is that the use of such an approach is an intrinsic feature of complex systems, be they artificial or natural. System Control: Various control schemes used in complex engineering systems are also found in various aspects of biological systems. Feedforward control and feedback control are two major control schemes, both of which are found almost ubiquitously in biological systems. Feedforward control is an open-loop control in which a set of predefined reaction sequences is triggered by a certain stimulus. Feedback is a sophisticated control system that closes the loop of the signal circuits to attain the desired control of the system. A negative feedback system detects the difference between desired output and actual output and compensates for such difference by modulating the input. While there are feedforward control methods, feedback control is more sophisticated and ensures proper control of the system and it can be used with feedforward control. It is one of the most widely used methods in engineering systems to increase the stability and robustness of the system. Redundancy: Redundancy is a widely used method to improve the system’s robustness against damage to its components by using multiple pathways to accomplish the function. Duplicated genes and genes with similar functions are basic examples of redundancy. There is also circuitlevel redundancy, such as multiple pathways of signal transduction and metabolic circuits that can be functionally complementary under different conditions. Modular Design: Modular design prevents damage from spreading limitlessly, and also improves ease of evolutionary upgrading of some of the 2 Discussion of similarity between complexity of engineering and biological systems as described in this section was first made, as far as the author is aware, by John Doyle at Caltech. 17

Systems Biology: Toward System-level Understanding of Biological Systems

Feedforward control








Feedback control



Figure 1.2 Feedforward control and feedback control

components. At the same time, a multi-functional module can help overcome system failure in a critical part by using modules in other less critical parts. Cellular systems are typical examples of modular systems. Structural Stability: Some gene regulatory circuits are built to be stable for a broad range of parameter variations and genetic polymorphisms. Such circuits often incorporate multiple attractors, each of which corresponds to functional state of the circuit; thus its functions are maintained against change in parameters and genetic polymorphisms. It is not clear whether such engineering wisdom is also the case in biological systems. However, the hypothesis is that such features are somewhat universal in all complex systems. It is conceivable that there are certain differences due to the nature of the system it is built upon, as well as the difference between engineering systems that are designed to exhibit certain functions and natural systems that have reproduction as a single goal where all functions are only evaluated in an integrated effect. Nevertheless, it is worth investigating the univerality of principles. And, if there are differences, what are they? The rest of the section focuses on how three principles of robustness exist also in biological systems. Of course, not all biological systems are robust, and it is important to know which parts of the systems are not robust and why. However, for this particular chapter, we will focus on robustness of biological systems, because it is one of the most interesting issues that we wish to understand. Control The use of explicit control scheme is an effective approach to improving robustness. Feedforward control and feedback control are two major methods of system control (Figure 1.2). Feedforward control is an open-loop control in which a sequence of predefined actions is triggered by a certain stimulus. This control method 18

Hiroaki Kitano

is the simplest method that works when possible situations and countermeasures are highly predictable. Feedback control, such as negative feedback, is a sophisticated control method widely used in engineering. It feeds back the sign-inverted error between the desired value and the actual value to the input, then the input signal is modulated proportional to the amount of error. In its basic form, it acts to minimize the output error value. Feedback plays a major role in various aspects of biological processes, such as E. coli chemotaxis and heat shock response, circadian rhythms, cell cycle, and various aspects of development. The most typical example is the integral feedback circuits involved in bacterial chemotaxis. Bacteria demonstrates robust adaptation to a broad range of chemical attractant concentrations, and so can always sense changes in chemical concentration to determine its behavior. This is accomplished by a circuit that involves a closed-loop feedback circuit (Alon et al., 1999; Barkai and Leibler, 1997). As shown in Figure 1.3, ligands that are involved in chemotaxis bind to a specific receptor MCP that forms a stable complex with CheA and CheW. CheA phosphorylates CheB and CheY. Phosphorylated CheB demethylates the MCP complex, and phosphorylated CheY triggers tumbling behavior. It was shown through experiments and simulation studies that this forms a feedback circuit which enables adaptation to changes in ligand concentration. Specifically, for any sudden change in the ligand concentration, the average activity level that is characterized by the tumbling frequency quickly converges to the steady-state value. This means that the system only detects acute changes of the ligand concentration that can be exploited to determine tumbling frequency, but is insensitive to the absolute value of ligand concentration. Therefore, the system can detect and control its behavior to move to a high attractant concentration area in the field regardless of the absolute concentration level without saturating its sensory system. Detailed analysis revealed that this circuit functions as an integral feedback (Yi et al., 2000) — the most typical automatic control strategy. In bacteria, there are many examples of sophisticated control embedded in the system. The circuit that copes with heat shock, for example, is a beautiful example of the combined use of feedforward control and feedback control (Figure 1.4). Upon heat shock, proteins in E. coli can no longer maintain their normal folding structures. The goal of the control system is to repair misfolding proteins by activating a heat shock protein (hsp), or to dissociate misfolding proteins by protease. As soon as heat shock is imposed, a quick translational modulation facilitates the production of σ 32 factor by affecting the three-dimensional structure of rpoH mRNA that encodes σ 32 . This leads to the formation of σ 32 -RNAP holo-enzyme that activates hsp that repair misfolded proteins. This process is feedforward control that pre-encodes the relationship between heat shock and proper course of reactions. In this process, there is no detection of misfolded pro19

Systems Biology: Toward System-level Understanding of Biological Systems








CheW CheA

CheB P


CheY CheZ

Figure 1.3 Bacterial chemotaxis related feedback loop

teins to adjust the translational activity of σ 32 . Independently, DnaK and DnaJ detect misfolded proteins and release σ 32 factor, that has been bound with DnaK and DnaJ. Free σ 32 activates transcription of hsp, so that misfolded proteins are repaired. This process is negative feedback control, because the level of misfolded proteins is monitored and it controls the activity of σ 32 factor. Another example demonstrating the critical role of the feedback system is seen in growth control of human cells. Growth control is one of the most critical parts of cellular functions. The feedback circuit involved in p53 presents a clear example of how feedback is used (Figure 1.5). When DNA is damaged, DNA-dependent kinase DNA-PK is activated. Also, ATM is phosphorylated, which makes ATM itself in an active state and promotes phosphorylation of the specific locus of the p53 protein. When this locus is phosphorylated, p53 no longer forms a complex with MDM2, and escapes from dissociation. The phosphorylation locus depends on what kind of stress is imposed on DNA. Under a certain stress, phosphorylation takes place at the Ser15 site of p53, and promotes transcription of p21 that eventually causes G1 arrest. In other cases, it promotes activation of apoptosis inducing genes, such as pig-3, and results in apoptosis. For those cells that entered G1 arrest, DNA-PK and ATM activity are lost as soon as DNA is repaired. The loss of DNA-PK and ATM activity decreases phosphorylation of p53, so p53 will bind with MDM2 and dissolve. Without phosphorylation, the p53 protein promotes mdm-2 transcription. It is interesting to know that mdm-2 protein forms a complex to deac20

Hiroaki Kitano

Heat Shock Normal Protein

dnaK dnaJ grpE GroES GroEL

Misfolded Protein

dnaK dnaJ grpE GroES GroEL

σE σ 70

σ 32 dnaK dnaJ grpE



σ 32 E σ 32

Figure 1.4 Heat shock response with feedforward and feedback control

tivate the p53 protein. This is another negative feedback loop embedded in this system. Redundancy Redundancy also plays an important role in attaining robustness of the system, and is critical for coping with accidental damage to components of the system. For example, the four independent hydraulic control systems in a Boeing 747 render the systems functionally normally even if one or two of them are damaged. In aircraft, control systems and engines are designed to have a high level of redundancy. In a cellular system, signal transduction and cell cycle are equivalent to control systems and engines. A typical signal transduction pathway is the MAP kinase cascade. The MAP kinase cascade involves extensive cross talk among collateral pathways. Even if one of these pathways is disabled due to mutation or other causes, the function of the MAP kinase pathway can be maintained because other pathways still transduce the signal (Figure 1.6). Cell cycle is the essential process of cellular activity. For example, in the yeast cell cycle, the Cln and Clb families play a dominant role in the progress of the cell cycle. They bind with Cdc28 kinase to form Cdk complex. Cln is redundant because knock-out of up to two of three Cln (Cln1, Cln2, Cln3) does not affect the cell cycle; all three Cln have to be knocked out to stop the cell cycle. Six Clb have very similar features, and may have originated in gene duplication. No single loss-of-function mutant of any of the six Clb affects growth of the yeast cell. The double mutants of CLB1 21

Systems Biology: Toward System-level Understanding of Biological Systems

p53 ATM









pig-3, etc.






G1 arrest

p DNA repair during G1 arrest

DNA damage Figure 1.5 p53 related feedback loop

and CLB2, as well as CLB2 and CLB3s are lethal, but other double mutant combinations do not affect phenotype. It is reasonable that the basic mechanism of the cell cycle has evolved to be redundant, thus robust against various perturbations. Redundancy can be exploited to cope with uncertainty involved in stochastic processes. McAdams and Arkin argued that duplication of genes and the existence of homologous genes improve reliability so that transcription of genes can be carried out even when only a small number of transcription factors are available (McAdams and Arkin, 1999). The use of a positive feedback loop to autoregulate a gene to maintain its own expression level is an effective means of ensuring the trigger is not lost in the noise. Although its functional implication has not been sufficiently investigated, an analysis of MAP kinase cascade revealed that it utilizes nonlinear properties intrinsic in each step of the cascade and positive feedback to constitute a stable all-or-none switch (Ferrell and Machleder, 1998). In the broader sense, the existence of metabolic pathways that can alternatively function to sustain cellular growth with changing environment can be viewed as redundancy. Bacteria is known to switch metabolic pathways if deprived of one type of nutrition, and to use other types of nutrition that are available. Theoretical analysis combined with experimental data indicate that different pathways are used to attain essentially the 22

Hiroaki Kitano

Raf, Mos









Transcription Figure 1.6 Redundancy in MAP kinase cascade

same objective function (Edward et al., 2001). Once we understand the stability and robustness of the system, we should be able to understand how to control and transform cells. We will then be ready to address such questions as how to transform cells that are malfunctioning into normal cells, how to predict disease risk, and how to preemptively treat potential diseases. Modular Design Modular design is a critical aspect of the robustness: it ensures that damage in one part of the system does not spread to the entire system. It may also ensure efficient reconfiguration throughout the evolutionary process to acquire new features. The cellular structure of the multicellular organism is a clear example. It physically partitions the structure so that the entire system does not collapse due to local damage. Gene regulatory circuits are considered to entail a certain level of modularity. Even if part of the circuit is disrupted due to mutation or injection of chemicals, it does not necessary affect other parts of the circuit. For example, mutation in p53 may destroy the cell cycle check point system that leads to cancer. However, it does not destroy metabolic pathways, so the cells continue to proliferate. How and why such modularity is maintained is not well understood at present. Modularity reflects hierarchical organization of the system that can be viewed as follows: Component: An elementary unit of the system. In electronics, transistors, capacitors, and resistors are components. In biological systems, genes and proteins, which are transcriptional products, are components. Device: A minimum unit of the functional assembly. NAND gates and


Systems Biology: Toward System-level Understanding of Biological Systems

flip-flops are examples of devices3 . Transcription complexes and replication complexes are examples of devices. Some signal transduction circuits may be considered as devices. Module: A large cluster of devices. CPU, memory, and amplifiers are modules. In biological systems, organella and gene regulatory circuits for the cell cycle are examples of modules. System: A top-level assembly of modules. Depending on the viewpoint, a cell or entire animal can be considered as a system. In engineering wisdom, each low-level module should be sufficiently self-contained and encapsulated so that changes in higher-level structure do not affect internal dynamics of the lower-level module. Whether is this also the case for biological systems and how it can be accomplished are of major interest from a system perspective. Structural Stability Some circuits may, after various disturbances to the state of the system, resume as one of multiple attractors (points or periodic). Often, feedback loops play a major role in making this possible. However, feedback does not explicitly control the state of the circuit in tracking or adapting to stimuli. Rather, dynamics of the circuit exhibit certain functions that are used in the larger sub-systems. The most well understood example is seen in one of the simplest organisms, lambda phage (McAdams and Shapiro, 1995). Lambda phage exploits the feedback mechanism to stabilize the committed state and to enable switching of its pathways. When lambda phage infects E. coli, it chooses one of two pathways: lysogeny and lysis. While a stochastic process is involved in the early stage of commitment, two positive and negative feedback loops involving CI and Cro play a critical role in stable maintenance of the committed decision. In this case, whether to maintain feedback or not is determined by the amount of activator binding to the O R region, and the activator itself cuts off feedback if the amount exceeds a certain level. This is an interesting molecular switch that is not found elsewhere. Overall, the concentration mechanism of Cro is maintained at a certain level using positive feedback and negative feedback. It was reported that the fundamental properties of the lambda phage switch circuit are not affected even if the sequence of O R binding sites is altered (Little et al., 1999). This indicates that properties of the lambda phage decision circuit are intrinsic to the multiple feedback circuit, not specific parametric features of the elements, such as binding sites. Relative independence from specific parameters is an important fea3 In electronics, “device” means transistors and other materials mentioned in “components.” NAND gates and flip-flops are recognized as minimum units of the circuit. 24

Hiroaki Kitano

ture of a robust system. Recent computational studies report that circuits that are robust against a broad range of parameter variations are found in Xenopus cell cycle (Morohashi et al., unpublished) and body segment formation (von Dassow et al., 2000). Using the simulation of parasegment formation of Drosophila, it was found that some parameters in the circuit accountable for pattern formation are tolerant to major parameter variations. This strongly suggests that the structure of the circuit that is dominantly responsible for pattern formation rather than specific parameter values (von Dassow et al., 2000). Such circuit features of structural stability also play important roles in development. A recent review article (Freeman, 2000) elucidates some interesting cases of feedback circuits that play a dominant role in the development process. Such cases include temporal arrangement of signaling in the JAK/STAT signaling pathway, pattern formation in Drosophila involving Ubx and Dpp, maintenance of patterns of expression for sonic hedgehog (Shh) that forms ZPA and Fgf, forming AER in limb development, etc. In these examples, structure of circuits play the dominant role rather than specific set of parameters. THE SYSTEOME PROJECT

In order to promote scientific research of systems biology, it is critically important to create a comprehensive data resource that describes systems’ features, as does the human genome project. This is an enormous challenge, and it requires significant efforts far beyond the capability of any single research group. Therefore, the author proposes “The Systeome Project” as a grand challenge in the area of systems biology. Systeome is an assembly of system profiles for all genetic variations and environmental stimuli responses. A system profile comprises a set of information on the properties of the system that includes the structure of the system and its behaviors, analysis results such as phase portfolio, and bifurcation diagrams. The structure of the system includes the structure of gene and metabolic networks and its associated constants, physical structures and their properties. Systeome is different from a simple cascade map, because it assumes active and dynamic simulations and profiling of various system statuses, not a static entity. The author suggests that the project be established for comprehensive efforts for profiling the Systeome of human, mouse, Drosophila, C. elegans, and yeast. The goal of the Human Systeome Project is defined as “To complete a detailed and comprehensive simulation model of the human cell at an estimated error margin of 20 percent by the year 2020, and to finish identifying the system profile for all genetic variations, drug responses, and environmental stimuli by the year 2030.” Undoubtedly, this is an ambitious project, and requires several mile25

Systems Biology: Toward System-level Understanding of Biological Systems

Dynamics Information High resolution imaging Expression profile Protein interactions, etc.

Components Information

Basic Model Information Basic Structure

System Dynamics Information System Dynamics Analysis (bifurcation, phase portfolio, etc.)

Gene Network Model Metabolic Pathway Model Signal Transduction Model

Mutation Analysis


Drug Sensitivity Analysis

Individual Genetic Variations


Individual Sequence Variations (SNPs, etc.)


Individual Heterochromatin Variations

Individual Systeome

Figure 1.7 Genome, Proteome, and Systeome

stones and pilot projects leading to the final goal. Initial pilot projects can using yeast and C. elegans be set with a time frame of five or seven years after full-scale budget approval. The Human Systeome Project shall be commenced concurrently with such pilot projects. The impact of this project will be far-reaching. It will be a standard asset for biological research as well as providing fundamental diagnostics and prediction for a wide range of medical practices. The Systeome Project is expected to contribute to system-level understanding of life by providing exhaustive knowledge of system structures, dynamics, and their sensitivities against genetic variations and environmental stimuli. By using the system profile, it is expected that more precise medical diagnosis and treatment can be accomplished due to quantitative understanding of the metabolic state of the system. For example, a list of all possible feedback loops and their sensitivities, gain, and time delay should be obtained, to be used for drug design and clinical applications. The behaviors of feedback systems are often counterintuitive and often eliminate or compensate the effects of external stimuli. Understanding of complex circuit dynamics such as these will contribute to accurate prediction of the effects of medical treatments. The Systeome Project should maintain close links with genome and Proteome data, particularly with various individual genetic variations, including single nucleotide polymorphisms (SNPs). SNPs are a typical example of an attempt to understand the relationship between genetic variations and clinical observations. It is inevitable that in some cases the effects of SNPs are masked by a mechanism that compensates such variations. In this case, corresponding SNPs do not seem to affect the behavior of the cell. However, if such a compensation mechanism is disrupted by SNPs in a locus that constitutes 26

Hiroaki Kitano

the compensation mechanism, the effects of SNPs will show up directly in the cell’s behavior. In such a case, it will be observed that for certain groups of cells, SNPs affect phenotype, but for other groups SNPs do not seem to affect phenotype. While SNPs provide certain information on individual variations at the genetic level, they do not provide the quantitative status of mRNA and proteins. Many biological phenomena have a certain quantitative sensitivity. The cell cycle, for example, is expected to take place when cyclin synthesis and degradation rate are within a certain range. SNPs and other existing genetic analysis cannot provide insights into quantitative aspects of such phenomena. Scientifically, a detailed understanding of circuits and their dynamics will contribute to a deeper understanding of the biological systems, as already discussed elsewhere. Identification of metabolic and signal transduction circuits in various model systems provides an interesting opportunity to compare evolutionary conserved genetic information not only at the gene level, but also at the circuit level. Evolutionary conserved circuits will be an important concept that may be widely used in the study of gene and metabolic network behaviors. Several circuits that may be found in yeast and C. elegans may be used also in mouse and human, similar to the idea of homologue genes. Some of the feedback circuits, for example, may be so essential that they have been conserved through the course of evolution. At the same time, a certain circuit may be duplicated and a revised version is used for other parts of the system. As the Systeome Project progresses in various model systems, such comparative studies and homology searches at the circuit level will become possible. Many scientific opportunities will open up once the Systeome Project has commenced and its data is made available for scientific research. The Systeome Project will be a major commitment. However, it is indispensable for promoting systems biology as quickly as possible and for contributing to a better understanding of living systems and for medical practice. The Systeome Project involves the major engineering project of developing the measurement and software platform. The best way to proceed with this project is to initiate it as an international joint project on a scale comparable to the human genome project. IMPACTS OF SYSTEMS BIOLOGY

Combined with the Systeome Project and other efforts in medical application of genomics, systems biology may have major impacts on medical research and practice. In-depth knowledge of the dynamical state of cells and development of high-performance measurement systems will drastically change medical practice. 27

Systems Biology: Toward System-level Understanding of Biological Systems

First, the fast and precise measurement of an individual systeome will enable us to make precise assessment and simulation of disease risk, as well as detailed planning of countermeasures. Establishment of “preemptive molecular medicine” is one of the major applications of systems biology research. This means that patient models, or disease models, can be grounded on the cellular model, instead of being an empirical phenomenological model. Second, drug design and treatment procedure may change to reflect the precise system dynamics of each patient. Rather than rely on a single drug, there many be increasing use of system drugs, a group of drugs that cooperatively act to control the metabolic state of malfunctioning cells. The point of such a treatment is to minimize side-effects, while maintaining maximum efficacy in disease treatment. By specifically identifying a series of effector points of chemical agents, we may be able to control cell status much more effectively than current medical practice. Third, system-level understanding, especially simulation, control, and design capability, may lead to a totally new method of organ cloning. Just like engineers perform digital pre-assembly, we may be able to digitally pregrow organs for transplant. There will be a special incubation system that can monitor and control a growing organ inside the incubator. Currently, regenerative medicine is now being practiced, but it is limited to re-generation of relatively simple tissue systems such as skin. For growing more complex organs such as the heart and kidney, sophisticated growth monitoring and control are required. This is “closed-loop manufacturing,” where the growth process is monitored and data is fed back to control the biochemical status of the incubation system to guide the organ growth to the desired shape. There will be many more medical applications. The Systeome Project is perhaps the best way to accelerate progress in the technology of systemlevel biology. CONCLUSION

Systems biology is a new and emerging field in biology that aims at system-level understanding of biological systems. System-level understanding requires a range of new analysis techniques, measurement technologies, experimental methods, software tools, and new concepts for looking at biological systems. The work has just begun and there is a long way to go before we arrive at a deep understanding of biological systems. Nevertheless, the author believes that systems biology will be the dominant paradigm in biology, and many medical applications as well as scientific discoveries are expected.


Hiroaki Kitano


The author would like to thank members of the Kitano Symbiotic Systems Project (Shuichi Onami, Shugo Hamahashi, Koji Kyoda, Mineo Morohashi, John Doyle, Mel Simon, Hamid Bolouri, Tau-Mu Yi, Mark Borisuk, Michael Hucka, Andrew Finny, Herbert Sauro, Yoshi Kubota) for fruitful discussions that helped me to form the idea in this chapter. In particular, John and Mel have always been strong supporters of systems biology and sources of inspiration. Shin-ichirou Imai has always been a great collaborator, and it was he who guided me to the area of biology, while I was still absorbed in computer science and robotics. Mario Tokoro and Toshi Doi allowed me to work on biology despite its tenuous link with the ongoing business of Sony.


Systems Biology: Toward System-level Understanding of Biological Systems

References Akutsu, T., Miyano, S., and Kuhara, S. (1999). Identification of genetic networks from a small number of gene expression patterns under the Boolean network model. Proc. Pacific Symposium on Biocomputing ’99 pp.17–28. Alon, U., Surette, M.G., Barkai, N., and Leibler, S. (1999). Robustness in bacterial chemotaxis. Nature 397:168–171. Barkai, N. and Leibler, S. (1997). Robustness in simple biochemical networks. Nature 387:913–917. Borisuk, M. and Tyson, J. (1998). Bifurcation analysis of a model of mitotic control in frog eggs. Journal of Theoretical Biology 195:69–85. Brown, P.O. and Botstein, D. (1999). Exploring the new world of the genome with DNA microarrays. Nature Genetics 21:33–37. Cannon, W.B., (1933). The wisdom of the body, Norton, New York. The C. elegans Sequencing Consortium. (1998). Genome sequence of the nematode C. elegans: A platform for investigating biology. Science 282:2012–2018. DeRisi, J.L., Lyer, V.R., and Brown, P.O. (1997). Exploring the metabolic and genetic control of gene expression on a genomic scale. Science 278:680–686. D’haeseleer, P., Wen, X., Fuhrman, S., and Somogyi, R. (1999). Linear modeling of mRNA expression levels during CNS development and injury. Proc. Pacific Symposium on Biocomputing ’99 pp.41–52. Edward, J.S. and Palsson, B.O. (1999). Systems properties of the Haemophilus influenzae Rd metabolic genotype. Journal of Biological Chemistry 274:17410–17416. Edward, J.S., Ibarra, R., and Palsson, B.O. (2001). In silico predictions of Escherichia coli metabolic capabilities are consistent with experimental data. Nature Biotechnology 19(2):125–130.

Fell, D.A. (1996). Understanding the control of metabolism, Portland Press, London. Ferrell, J. and Machleder, E. (1998). The biochemical basis of an all-or-none cell fate switch in Xenopus ooctytes. Science 280:895–898. Freeman, M. (2000). Feedback control in intercellular signalling in development. Nature 408:313–319. Hamahashi, S. and Kitano, H. (1998). Simulation of fly embryogenesis. Proc. the 6th International Conference on Artificial Life pp.151–160. Hamahashi, S. and Kitano, H. (1999). Parameter optimization in hierarchical structure. Proc. the 5th European Conference on Artificial Life pp.467– 471. Heinrich, R. and Rapoport, T.A. (1974). A linear steady-state treatment of enzymatic chains. Eur. J. Biochem. 42:89–95. Hucka, M., Sauro, H., Finney, A., and Bolouri, H. (2000). An XML-based model description language for systems biology simulations. Working Draft, ERATO Kitano Project – CALTECH Group. Hucka, M., Finney, A., Sauro, H., Bolouri, H., Doyle, J., and Kitano, H. (2001). The ERATO Systems Biology Workbench: An integrated environment for multiscale and multitheoretic simulations in systems biology. Foundations of Systems Biology, The MIT Press, Cambridge. Ideker, T., Thorsson, V., and Karp, R. (2000). Discovery of regulatory interactions through perturbation: inference and experimental design. Proc. Pacific Symposium on Biocomputing 2000 pp.302–313. Imai, S. and Kitano, H. (1998). Heterochromatin island and their dynamic reorganization: A hypothesis for three distinctive features of cellular aging. Experimental Gerontology 33(6):555–570. Ito, T., Tashiro, K., Muta, S., Ozawa, R., Chiba, T., Nishizawa, M., Yamamoto, K., Kuhara, S., and Sakaki, Y. (2000). Toward a protein-protein interaction map of the budding yeast: A comprehensive system to examine two-hybrid interactions in all possible combinations between the yeast proteins. Proc. Natl. Acad. Sci. USA 97(3):1143–1147. Kacser, H. and Burns, J. A. (1973). The control of flux. Symp. Soc. Exp. Biol. 27:65–104. Kanehisa, M., and Goto, S. (2000). KEGG: Kyoto encyclopedia of gene and genomes. Nucleic Acids Res. 28:29–34. Karp, P., Paley, M., Pellegrini-Toole, A., Krummenacker, M. (1999). EcoCyc: Electronic encycropedia of E. coli genes and metabolism. Nucleic Acids Res. 27(1):55. 32


Kitano, H. (2000). Perspectives on systems biology. New Generation Computing 18(3):199–216. Kitano, H. and Imai, S. (1998). The two-process model of cellular aging. Experimental Gerontology 33(5):393–419. Kitano, H., Hamahashi, S., Takao, K., and Imai, S. (1997). Virtual biology laboratory: A new approach of computational biology. Proc. the 4th European Conference on Artificial Life pp.274–283. Kitano, H., Hamahashi, S., and Luke, S. (1998). The Perfect C. elegans Project: An initial report. Artificial Life 4:141–156. Kondo, S. and Asai, R. (1995). A reaction-diffusion wave on the skin of the marine angelfish Pomacanthus. Nature 376:765–768. Koza, J., Mydlowec, W., Lanza, G., Yu, J., and Keane, A. (2001). Automated reverse engineering of metabolic pathways from observed data by means of genetic programming. Foundations of Systems Biology, The MIT Press, Cambridge. Kyoda, K. and Kitano, H. (1999). Simulation of genetic interaction for Drosophila leg formation. Proc. Pacific Symposium on Biocomputing ’99 pp.77–89. Kyoda, K. and Kitano, H. (1999). A model of axis determination for the Drosophila wing disc. Proc. the 5th European Conference on Artificial Life pp.472–476. Kyoda, K., Muraki, M., and Kitano, H. (2000). Construction of a generalized simulator for multi-cellular organisms and its application to SMAD signal transduction. Proc. Pacific Symposium on Biocomputing 2000 pp.314–325. Kyoda, K., Morohashi, M., Onami, S. and Kitano, H. (2000). A gene network inference method from continuous-value gene expression data of wild-type and mutants. Genome Informatics 11:196–204. Lagally, E.T., Medintz, I., and Mathies, R.A. (2001). Single-molecule DNA amplication and analysis in an integrated microfluidic device. Anal. Chem. 73:565–570. Liang, S., Fuhrman, S., and Somogyi, R. (1999). REVEAL, a general reverse engineering algorithm for inference of genetic network architectures. Proc. Pacific Symposium on Biocomputing ’99 pp.18–29. Little, J.W., Shepley, D.P., and Wert, D.W. (1999). Robustness of a gene regulatory circuit. EMBO J. 18(15):4299–4307.



McAdams, H.H. and Arkin, A. (1999). It’s a noisy business! Genetic regulation at the nanomolar scale. Trends in Genetics 15(2):65–69. McAdams, H.H. and Arkin, A. (1998). Simulation of prokaryotic genetic circuits. Annu. Rev. Biophys. Biomol. Struct. 27:199–224. McAdams, H. and Shapiro, L. (1995). Circuit Simulation of genetic networks. Science 269:650–656. Mendes, P. and Kell, D.B. (1998). Non-linear optimization of biochemical pathways: Applications to metabolic engineering and parameter estimation. Bioinformatics 14(10):869–883. Michaels, G.S., Carr, D.B., Askenazi, M., Fuhrman, S., Wen, X., and Somogyi, R. (1998). Cluster analysis and data visualization of large-scale gene expression data. Proc. Pacific Symposium on Biocomputing’98 pp.42–53. Morohashi, M. and Kitano, H. (1998). A method for reconstructing genetic regulatory network for Drosophila eye formation. Proc. the 6th International Conference on Artificial Life pp.72–80. Morohashi, M. and Kitano, H. (1999). Identifying gene regulatory networks from time series expression data by in silico sampling and screening. Proc. the 5th European Conference on Artificial Life pp.477–486. Morohashi, M., Winn, A.E., Borisuk, M.T., Bolouri, H., Doyle, J., and Kitano, H. Robustness as a measure of plausibility in models of biochemical networks. Unpublished. Nagasaki, M., Onami, S., Miyano, S., and Kitano, H. (1999). Bio-Calculus: Its concept and molecular interaction. Genome Informatics 10:133–143. Okuno, G.H., Kyoda, K., Morohashi, M., and Kitano, H. (1999). An initial assessment of ERATO-1 Beowulf-class cluster. Proc. International Workshop on Parallel and Distributed Computing for Symbolic and Irregular Applications. Onami, S., Hamahashi, S., Nagasaki, M., Miyano, S., and Kitano, H. (2001). Automatic acquisition of cell lineage through 4D microscopy and analysis of early C. elegans embryogenesis. Foundations of Systems Biology, The MIT Press, Cambridge. Onami, S., Kyoda, K.M., Morohashi, M., and Kitano, H. (2001). The DBRF method for inferring a gene network from large-scale steady-state gene expression data mutants. Foundations of Systems Biology, The MIT Press, Cambridge. Reinitz, J., Mjolsness, E., and Sharp, D.H. (1995). Model for cooperative control of positional information in Drosophila by bicoid and maternal hunchback. J. Exp. Zoo. 271:47–56. 34


Savageau, M.A., Voit, E.O., and Irvine, D.H. (1987). Biochemical systems theory and metabolic control theory: 1. Fundamental similarities and differences. Mathematical Biosciences 86:127–145. Simpson, P., Roach, D., Woolley, A., Thorson, T., Johnston, R., Sensabaugh, G., and Mathies, G. (1998). High-throughput genetic analysis using microfabricated 96-sample capillary array electrophoresis microplates. Proc. Natl. Acad. Sci. USA 95:2256–2261. Schwikowski, B., Uetz, P., and Fields, S. (2000). A network of proteinprotein interactions in yeast. Nature Biotech. 18:1257–1261. Spellman, P.T., Sherlock, G., Zhang, M.Q., Iyer, V.R., Anders, K., Eisen, M., Brown, P.O., Botstein, D., and Futcher, B. (1998). Comprehensive identification of cell cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization. Molecular Biol. Cell. 9:3273–3297. Sulston, J.E. and Horvitz, H.R. (1997). Post-embryonic cell lineage of the nematode Caenorhabditis elegans. Dev. Biol. 56:110–156. Sulston, J.E., Schierenberg, E., White, J.G., and Thomson, J.N. (1983). The embryonic cell lineage of the nematode Caenohabditis elegans. Dev. Biol. 100:64–119. Tabara, H., Motohashi, T., and Kohara, Y. (1996). A multi-well version of in situ hybridization on whole mount embryos of Caenorhabditis elegans. Nucleic Acids Research 24:2119–2124. Tomita, M., Shimizu, K., Matsuzaki, Y., Miyoshi, F., Saito, K., Tanida, S., Yugi, K., Venter, C., and Hutchison, C. (1999). E-Cell: Software environment for whole cell simulation. Bioinformatics 15(1):72–84. Ueda, H. and Kitano, H. (1998). A generalized reaction-diffusion simulator for pattern formation in biological systems. Proc. the 6th International Conference on Artificial Life pp.462–466. Varma, A. and Palsson, B.O. (1994). Metabolic flux balancing: Basic concepts, scientific and practical use. Bio/Technology 12:994–998. von Bertalanffy, L. (1968). General System Theory, Braziler, New York. von Dassow, G., Meir, E., Munro, E.M., and Odell, G. (2000). The segment polarity network is a robust developmental module. Nature 406:188– 192. Watson, J. D. and Crick, F.H. (1953). Molecular structure of nucleic acids: A structure for deoxyribose Nucleic Acid. Nature 171:737–738. White, J.G., Southgate, E., Thomson, J.N., and Brenner, S. (1986). The structure of the nervous system of the nematode Caenorhabditis elegans. Phil. Trans. R. Soc. 314:1–340. 35


Wiener, N., (1948). Cybernetics or Control and Communication in the Animal and the Machine, The MIT Press, Cambridge. Yasuda, T., Bannai, H., Onami, S., Miyano, S., and Kitano, H. (1999). Towards automatic construction of cell-lineage of C. elegans from Normarski DIC microscope images. Genome Informatics 10:144–154. Yi, T.-M., Huang, Y., Simon, M., and Doyle, J. (2000). Robust perfect adaptation in bacterial chemotaxis through integral feedback control. Proc. Natl. Acad. Sci. USA 97(9):4649–4653.



Part I Advanced Measurement Systems


Automatic Acquisition of Cell Lineage through 4D Microscopy and Analysis of Early C. elegans Embryogenesis Shuichi Onami, Shugo Hamahashi, Masao Nagasaki, Satoru Miyano, and Hiroaki Kitano

Cell lineage analysis is an important technique for studying the development of multicellular organisms. We have developed a system that automatically acquires cell lineages of C. elegans from the 1-cell stage up to approximately the 30-cell stage. The system utilizes a set of 4D Nomarski DIC microscope images of C. elegans embryo consisting of more than 50 focal plane images at each minute for about 2 hours. The system detects the region of cell nucleus in each of the images, and makes 3D nucleus regions, each of which is a complete set of nucleus regions that represent the same nucleus at the same time point. Each pair of 3D nucleus regions is then connected, if they represent the same nucleus and their time points are consecutive, and the cell lineage is created based on these connections. The resulting cell lineage consists of the three-dimensional positions of nuclei at each time point and their lineage. Encouraged by the performance of our system, we have started systematic cell lineage analysis of C. elegans, which will produce a large amount of quantitative data essential for system-level understanding of C. elegans embryogenesis. INTRODUCTION

In the last few decades, biology has been mainly focusing on identifying components that make up the living system. Today, as a result of success in molecular biology and genomics, thousands of genes have been identified, so the focus of biology is now moving toward the next step, understanding how those components work as a whole system. The ultimate goal of this step is the computer simulation of life, that is, reconstruction of living systems on the computer. However, it is still difficult to perform reliable computer simulation even of a single cell. The first reason for this difficulty is the lack of biological knowledge. Based on the genomic sequence, the number of human genes was predicted as approximately 26,000 (Venter et al., 2001), and about 4300 genes were predicted even for Eschericia coli, a single-cellular prokaryote (Barrick

et al., 1994). However, with some exceptions, information on the functions of those genes is quite limited. In order to determine those functions efficiently, automation of biological experiments is necessary. Such automation, following on from the automatic DNA sequencer, the DNA microarray system, etc., will greatly increase the quality of computer simulation. The second reason is the lack of quantitative information. Historically, biology has been mainly accumulating qualitative information, such as “the expression of gene increases” and “the nucleus moves to the anterior.” However, for computer simulation, quantitative information is necessary, such as “the expression of gene increases at v ng/s,” and “the position of the nucleus is (x, y, z).” To obtain such quantitative information, precisely controlled analytical instruments need to be developed. The third reason is the immaturity of modeling technology and simulation technology. Several software packages have been developed for biological computer simulation (Mendes, 1993; Morton-Firth and Bray, 1998; Tomita et al., 1999; Shaff and Loew, 1999; Kyoda et al., 2000). These efforts have greatly improved modeling and simulation technology for simple biological processes, such as reactions among free molecules, and so the accuracy and reliability of computer simulation have been greatly increased for single-cellular organisms and individual cells. However, almost no technology has been developed for more complicated biological processes, such as sub-cellular localization of molecules and organelles, cell division, and three-dimensional positioning of cells. These technologies are essential for reliable simulation of multicellular organisms. This chapter reviews our automatic cell lineage acquisition system, which is one of our approaches we have developed for computer simulation of C. elegans. The system automates biological experiments and produces quantitative data. The end of this chapter briefly reviews our other approaches, which are improving modeling and simulation technology, and then briefly overviews our approaches as a whole. THE NEMATODE, C. ELEGANS

There are good introductions to C. elegans in other literatures (Wood et al., 1988; Riddle et al., 1997), so a detailed introduction of this organism is omitted. Briefly, C. elegans is the simplest multicellular organism that has been most extensively analyzed in molecular and developmental biology. This organism is also the first multicellular organism whose genome sequence has been completely identified (The C. elegans Sequencing Consortium, 1998), and is leading the other multicellular experimental organisms in post genome sequencing analysis, such as functional genomics and proteomics. Thus, C. elegans is expected to be the first multicellular organism whose life is fully reconstructed on the computer.


Shuichi Onami, et al.

Figure 2.1 Cell lineage. When the fertilized egg undergoes a series of cell divisions shown on the left, the cell lineage is described as shown on the right. In the cell lineage, the vertical axis represents the time and the horizontal axis represents the direction of division (left-right and anterior-posterior).

Figure 2.2 The complete cell lineage of C. elegans (Sulston et al., 1983).


Generally, a multicellular organism is a mass of cells that are generated from a single cell – i.e. the fertilized egg – through successive cell divisions. Each cell division is a process whereby a single mother cell produces a pair of daughter cells. Cell lineage is a tree-like description of such mother-daughter relationships starting from the fertilized egg (in a wide sense, starting from a specific cell) (Moody, 1999) (Figure 2.1). It usually includes information on the timing and the direction of each cell division. The complete cell lineage – from the fertilized egg to the adult – has been identified for several simple multicellular organisms, such as C. elegans (Sulston et al., 1983) and Halocynthia roretzi (Nishida, 1987)(Figure 2.2). The most typical application of cell lineage is gene function analysis


Automatic Acquisition of Cell Lineage

Figure 2.3 Comparison of cell lineage between wild type and mutant animals. When the wild type and the mutant cell lineages are described as in this figure, the mutated gene plays some roles in the differentiation of the two daughter cells produced at the first cell division

by comparing cell lineages among wild type and mutant animals (Figure 2.3). Through such analysis, many gene functions have been uncovered. So, cell lineage analysis is an important technique for studying the development of multicellular organisms, as well as in situ hybridization, immunohistochemistry, and GFP-fusion gene expression. HISTORY OF CELL LINEAGE ANALYSIS PROCEDURE

This section reviews the history of the cell lineage analysis procedure, focusing on the procedure for C. elegans. But with some difference in the dates, the history is almost the same in other animals. The entire cell lineage of C. elegans was reported by Sulston et al. in 1983 (Sulston et al., 1983). In this work, they used a rather primitive procedure whereby they directly observed the animal through a Nomarski DIC microscope and sketched it. A Nomarski DIC microscope visualizes subtle differences of thickness and refraction index in the light path, and has an advantage that the intra-cellular structure of living transparent cells can be studied without staining (Spector et al., 1998). Through this microscope, moving the focal plane up and down, Sulston et al. observed and sketched a 14-hour process of C. elegans embryogenesis, from fertilization to hatching (Figure 2.4), which must have been quite laborious. The four-dimensional microscope imaging system (4D microscope), developed by Hird et al. in 1993, greatly reduced the laboriousness of cell lineage analysis (Hird and White, 1993). By controlling the focusing device and the camera, the system automatically captures microscope images of different focal planes, and repeats the process with a given interval. This system obtains a set of microscope images that contain the 3D structure information of an embryo at different time points starting 42

Shuichi Onami, et al.

Figure 2.4 Nomarski DIC microscope images of different focal planes. Nomarski DIC microscope images of a 2-cell stage embryo are shown. Moving the focal plane up and down, the 3D structure of the embryo can be recognized.

from fertilization (Figure 2.5). Then, those images are closely analyzed to derive the cell lineage. A GUI supporting tool, developed by Schnabel et al. in 1997, further reduced the laboriousness of cell lineage analysis (Schnabel et al., 1997). As is reviewed above, cell lineage analysis has become much easier than that in Sulston’s era, but it is still quite laborious. The number of mutants whose cell lineage is identified, is quite small compared with the number of mutants whose responsible gene is identified and sequenced. AUTOMATIC CELL LINEAGE ACQUISITION

We are developing a system that automatically acquires cell lineages of C. elegans (Yasuda et al., 1999). The latest version of our system has the ability to acquire the cell lineage from the 1-cell stage up to approximately the 30-cell stage (Hamahashi et al., unpublished). This section reviews the process of our system. The system utilizes a set of 4D Nomarski DIC microscope images to extract the cell lineage (Figure 2.5). The 4D microscope system is able to capture more than 50 images per minute, changing the focal plane position by 0.5 µm for each image. With this system, a set of multi-focal plane images of a C. elegans embryo is captured every minute for about


Automatic Acquisition of Cell Lineage

Figure 2.5 4D microscope images.

Figure 2.6 Example of an nucleus detection filter.

2 hours. Since the height of the embryo is about 25 µm, each multi-focal plane image includes all 3D structure information of the embryo at the corresponding time point. The system then processes each of the images captured in the previous step, and detects the regions of cell nucleus in the image (Figure 2.6). In the Nomarski microscope image, the region of cytoplasm looks bumpy as a result of the existing organelles, such as lysosome and mitochondria. On the other hand, the nucleus region, without those organelles, looks smooth. We found that several basic image-processing filters (i.e. Kirsch’s edge detection filter (Kirsch, 1971), entropy filter (Jahne et al., 1999), etc.) efficiently detect those nucleus regions (Yasuda et al., 1999; Hamahashi et 44

Shuichi Onami, et al.

Figure 2.7 Detected nucleus regions. Each of the detected nucleus regions is enclosed by a white line.

al., unpublished). Several new filters applicable to this nucleus detection were also developed (Yasuda et al., 1999; Hamahashi et al., unpublished). By appropriately combining those filters, we established an excellent nucleus detection algorithm (Hamahashi et al., unpublished). With this algorithm, non-error nucleus detection is carried out from the 1-cell stage to about the 30-cell stage (Figure 2.7). In wild type embryo, the diameter of a nucleus is about 7 µm at the 2cell stage and 4.5 µm at the 20-cell stage, whereas our system captures microscope images every 0.5 µm of focal plane position. Therefore, at every time point, a nucleus is detected on several different focal planes. In the next step, the system makes 3D nucleus regions, each of which is a complete set of nucleus regions that represent the same nucleus at the same time point. Then, the system connects each pair of 3D nucleus regions, if they represent the same nucleus and their time points are consecutive. In the previous two steps, a pair of nucleus regions is recognized as representing the same nucleus, when one nucleus region overlaps the other either on the same focal plane at the next time point or on the next focal plane at the same time point. As the result, the lineage of 3D nucleus regions is recognized through out the entire period of the 4D microscope images. Finally, the cell lineage is created based on the above 3D nucleus region lineage. The system calculates the centroid position of each 3D nucleus region, and outputs those centroid positions and their lineage (Figure 2.8). This section briefly reviews the process of our cell lineage detection system. The current system utilizes our Beowulf PC cluster (Okuno et al., 2000), made up of 32 PCs, to execute all the above processes except 4D microscope image recording, and within 9 hours, can deduce the cell lineage up to the 30- to 40-cell stage after setting the 4D microscope images 45

Automatic Acquisition of Cell Lineage

Figure 2.8 Text data for C. elegans cell lineage.

(Hamahashi et al., unpublished). We also developed a software package that three-dimensionally visualizes the resulting lineage data (Hamahashi et al., unpublished), which may help three-dimensional understanding of nucleus movement and division (Figure 2.9). Moreover, with this package, lineages of two different individuals – e.g., wild type and mutant – can be visualized on the same screen. ADVANTAGES OF AUTOMATIC CELL LINEAGE ACQUISITION SYSTEM

As noted in the previous section, we have successfully developed a highperformance automatic cell lineage acquisition system. The advantages of this system are outlined below. The most significant advantage of this system is automation, as can easily be imagined from the contribution of the automatic DNA sequencer to biology. The required human effort for our system is almost the same as that of the DNA sequencer. The processing time of 9 hours is almost the same as that of the sequencer in its early days. With this system, large scale and systematic cell lineage analysis is made possible. The second advantage is quantitative data production. The threedimensional nucleus position at each time point which the system outputs is quantitative data, which is essenaital for simulation studies, especially when simulation models are developed and simulation results are analyzed. Our system can be applied to many individual animals, wild types and mutants, and the resulting data will greatly improve the accuracy of computer simulation. The third advantage is the reproducibility of the results. When cell lin46

Shuichi Onami, et al.

Figure 2.9 Three-dimensional view of a C. elegans cell lineage. The centroid positions of 3D nucleus regions are traced from 1-cell to 19-cell stage. The white circles represent the centroid positions at the viewing time point. On this viewer, it is possible to freely change the viewing time point forward and backward, and also rotate the viewing point three-dimensionally.

eages are manually analyzed, the resulting cell lineage is unreproducible. For example, the definitions of nucleus position and cell division time may vary depending on who made the analysis and, even if the same people made it, when it was done. In our system, such definitions are exactly the same through all individual measurements and the results are completely reproducible. The results are thus suitable for statistical analysis, such as calculating the mean, variance, standard error, etc. Fourthly, the system is applicable to other organisms. The basis of the system is an image-processing algorithm for Nomarski microscope images. Thus, in principle, the system is applicable to all transparent cells and embryos. Future application to other organisms, such as Halocynthia roretzi, mouse, is promising. Finally, the system offers complementarity of cell lineage data. As a result of success in molecular biology and genomics, a variety of largescale analyses are currently undertaken, such as DNA mircoarray, protein chip, and systematic in situ hybridization. However, combinations of those analyses are not so fruitful since they all measure the same object – gene expression level. Cell lineage data is quite complementary to gene expression data, therefore, the combination of our cell lineage analysis with gene expression analyses will provide useful information for biology.


Automatic Acquisition of Cell Lineage

Figure 2.10 Systematic cell lineage analysis of C. elegans.


As described in this chapter, the current version of our system extracts a C. elegans cell lineage of up to the 30-cell stage in 9 hours. The biggest challenge in the future system development is, of course, to extend the applicable embryonic period, up to the 100-cell stage, 200cell stage, and beyond. The current limit of the applicable period is imposed by the performances of the 4D microscope system, such as the speed of the z-axis driving motor and the image-capturing period of the CCD camera. The performance of these devices is rapidly being improved, so such device-dependent limit will likely be overcome in the near future. The limit of the current algorithm may be around the 60-cell stage. For the later stages, an improved algorithm will be required. Nucleus detection is quite difficult even for humans after the 100-cell stage, so for later stages, GFP-labeling of nucleus or other nucleus labeling techniques may be necessary. Shortening the processing time is another important challenge, but the solution seems relatively easy. Dramatic improvement of CPU speed will greatly shorten the processing time of our system. SYSTEMATIC CELL LINEAGE ANALYSIS

Encouraged by the performance of the current cell lineage acquisition system, we have started systematic cell lineage analysis of C. elegans embryo (Figure 2.10). As the first step, we are currently accumulating many wild type cell lineages in order to establish the standard wild type cell lineage, which describes the mean value of nucleus position at each time point together with some statistical data, such as the variance, error distribution, etc. As well as wild type animals, we are also analyzing cell lineages for many 48

Shuichi Onami, et al.

mutants that are already known to play important roles in early embryogenesis. By analyzing the results, we will confirm and also improve the current understanding of early embryogenesis. In C. elegans, there is a quite well organized mutant-stocking system (Caenorhabditis Genetic Center 1 ). In addition, several whole genome knock-outing projects are being undertaken either by efficient mutagenesis (Gengyo-Ando and Mitani, 2000) or RNAi (Fraser et al., 2000), taking advantage of the complete genome sequence data (The C. elegans Sequencing Consortium, 1998). We are planning to start a systematic cell lineage analysis for those knock-out animals in future. The resulting data, together with the systematic gene expression data (Tabara et al., 1996), will provide useful information for the complete understanding of C. elegans. COMPUTER SIMULATION OF C. ELEGANS

Our cell lineage system can produce a large amount of quantitative data, which is useful for computer simulation. To achieve computer simulation of C. elegans embryogenesis, the authors are also running several other closely related projects, as outlined below. The quality of computer simulation is largely dependent on the simulation model, thus the model construction is an important process in simulation studies. To help this process, we are developing gene regulatory network inference methods. An efficient method has been developed for large-scale gene expression data, such as DNA microarray data (Kyoda et al., 2000). Currently, we are developing a sophisticated gene network modeling scheme based on this method, and are also trying to develop a gene network inference method that utilizes the cell lineage information. For improving the technology of biological computer simulation, we are developing a model description language for biological computer simulation (Nagasaki et al., 1999), named bio-calculus. The language will be able to describe a variety of biological processes observed in multicellular organisms, such as sub-cellular positioning of molecules and organelles, cell division, and three-dimensional positioning of cells. We are also developing several software packages so that a variety of biological models described using the language can be executed (Nagasaki et al., 1999). Currently, a very early period of C. elegans embryo is being modeled and simulated (Nagasaki et al., unpublished) in order to improve the applicability of the language and its software packages (Figure 2.11, 2.12). In future, utilizing our cell lineage information, we will gradually refine our C. elegans model and extend the target period, to improve the modeling and simulation technology further.

1 49

Automatic Acquisition of Cell Lineage

Figure 2.11 Pronucleus movement of C. elegans embryo. a)–c) Nomarski DIC microscope images of very early C. elegans embryo just after fertilization. The anterior is left. The oocyte pronucleus (left) and the sperm pronucleus (right) move toward each other and finally meet in the posterior hemisphere. The movement of the sperm pronucleus mainly depends on microtubules (MTs) (Hird and White, 1993). d)–f) Confocal microscope images of C. elegans embryo stained with MT specific antibody. MTs are growing from the centrosomes on the sperm pronucleus.


This chapter reviewed our automatic cell lineage acquisition system. The system will produce a large amount of quantitative data, which is valuable for computer simulation, though the data are still insufficient for the complete C. elegans simulation. We must therefore keep developing new experimental technologies. It is hoped that all our approaches will functionally work together to enable us to achive the ultimate goal – the computer simulation of life. ACKNOWLEDGEMENT

The authors thank Nick Rhind for his cell lineage drawing.


Shuichi Onami, et al.

Figure 2.12 Computer simulation of MT dependent sperm pronucleus movement in C. elegans embryo. The small circle represents the sperm pronucleus and the white lines growing from the sperm pronucleus represents MTs growing from the centrosomes on the pronucleus.


Automatic Acquisition of Cell Lineage

References Barrick, D., Villanueba, K., Childs, J., Kalil, R., Schneider, T.D., Lawrence, C.E., Gold, L., and Stormo, G.D. (1994). Quantitative analysis of ribosome binding sites in E. coli. Nucleic Acids Res. 22:1287–1295. The C. elegans Sequencing Consortium. (1998). Genome sequence of the nematode C. elegans. a platform for investigating biology. Science 282:2012–2018. Fraser, A., Kamath, R.S., Zipperlen, P., Martinez-Campos, M., Sohrmann, M., and Ahringer, J. (2000). Functional genomic analysis of C. elegans chromosome I by systematic RNA interference. Nature 408:325–330. Gengyo-Ando, K. and Mitani, S. (2000). Characterization of mutations induced by ethyl methanesulfonate, uv, and trymethylpsoralen in the nematode Caenorhabditis elegans. Biochem. Biophys. Res. Comm. 269:64–69. Hamahashi, S., Onami, S., and Kitano, H. Unpublished. Hird, S. and White, J.G. (1993). Cortical and cytoplasmic flow polarity in early embryonic cells of Caenorhabditis elegans. J. Cell Biol. 121:1343–1355. Jahne, B., Haussecker, H., and Geissler, P. editors. (1999). Handbook of computer vision and applications. Academic Press. Kirsch, R. (1971). Computer determination of the constituent structure of biological images. Comput. Biomed. Res. 4:315–328. Kyoda, K.M., Morohashi, M., Onami, S., and Kitano, H. (2000). A gene network inference method from continuous-value gene expression data of wild-type and mutants. Genome Informatics, 11:196–204. Kyoda, K.M., Muraki, M., and Kitano, H. (2000). Construction of a generalized simulator for multi-cellular organisms and its application to smad signal transduction. Proc. Pacific Symposium on Biocomputing 2000 pp.317–328. Mendes, P. (1993). Gepasi: A software package for modelling the dynamics, steady states and control of biochemical and other systems. Comput. Applc. Biosci. 9:563–571.

Moody, S.A., editor. (1999). Cell lineage and fate determination. Academic Press. Morton-Firth, C.J., and Bray, D. (1998). Predicting temporal fluctuations in an intracellular signalling pathway. J. Theor. Biol. 192:117–128. Nagasaki, M., Miyano, S., Onami, S., and Kitano, H. Unpublished. Nagasaki, M., Onami, S., Miyano, S., and Kitano, H. (1999). Bio-calculus: Its concept and molecular interaction. Genome Informatics 10:133–143. Nishida, H. (1987). Cell lineage analysis in ascidian embryos by intracellular injection of a tracer enzyme. III. up to the tissue restricted stage. Dev. Biol. 121:526–541. Okuno, H.G., Kyoda, K.M., Morohashi, M., and Kitano, H. (2000). Initial assessment of ERATO-1 beowulf-class cluster. Proc. Parallel and distributed computing for symbolic and irregular applications pp.372–383. Riddle, D.L., Blumenthal, T., Meyer, B.J., and Priess, J.R., editors. (1997). C. elegans II. Cold Spring Harbor Laboratory Press. Schaff, J. and Loew, L.M. (1999). The virtual cell. Proc. Pacific Symposium on Biocomputing ’99 pp.228–239. Schnabel, R., Hutter, H., Moerman, D., and Schnabel, H. (1997). Assessing normal embryogenesis in Caenorhabditis elegans using 4D microscope: variability of development and regional specification. Dev. Biol. 184:234– 265. Spector, D.L., Goldman, R.D., and Leinwand, L.A. editors. (1988). Cells - A Laboratory Manual. Cold Spring Harbor Laboratory Press. Sulston, J.E., Shierenberg, E., White, J.G., and Thomson, J.N. (1983). The embryonic cell lineage of the nematode Caenorhabditis elegans. Dev. Biol. 100:64–119. Tabara, H., Motohashi, T., and Kohara, Y. (1996). A multi-well version of in situ hybridization on whole mount embryos of Caenorhabditis elegans. Nucleic Acids Res. 24:2119–2124. Tomita, M., Hashimoto, K., Takahashi, K., Shimizu, T.S., Matsuzaki, Y., Miyoshi, F., Saito, F., Tanita, S., Yugi, K., Vender, J.C., and Hutchison, C.A. (1999). E-cell: software environment for whole-cell simulation. Bioinformatics 15:72–84. Venter, J.C., et al. (2001). The sequence of the human genome. Science 291:1304–1351. Wood, W.B. and the Community of C. elegans Researchers, editors. (1988). The nematode Caenorhabditis elegans. Cold Spring Harbor Laboratory. 54


Yasuda, T., Bannai, H., Onami, S., Miyano, S., and Kitano, H. (1999). Towards automatic construction of cell lineage of C. elegans from Nomarski DIC microscope images. Genome Informatics 10:144–154.



Part II Reverse Engineering and Data Mining from Gene Expression Data


The DBRF Method for Inferring a Gene Network from Large-Scale Steady-State Gene Expression Data Shuichi Onami, Koji M. Kyoda, Mineo Morohashi, and Hiroaki Kitano

Complete genome sequence has enabled whole-genome expression profiling and genome deletion projects, which are generating large-scale gene expression profiles corresponding to hundreds of deletion mutants. To obtain valuable information from those profiles is an important challenge in current biology. This chapter reviews the Difference-Based Regulation Finding (DBRF) method, which infers the underlying gene network from those profiles. The method 1) infers direct and indirect gene regulations by interpreting the difference of gene expression level between wild-type and mutant, and 2) eliminates the indirect regulations. One of the major characteristics of the method is its applicability to continuous-value expression data, whereas the other existing method can only deal with binary data. The performance of the method was evaluated using artificial gene networks by varying the network size, indegree of each gene, and the data characteristics (continuous-value or binary). The results showed that the method is superior to the other methods. The chapter also reviews the applicability of the DBRF method to real gene expression data. The method was applied to a set of yeast DNA microarray data which consisted of gene expression levels of 249 genes in each of single gene deletion mutants for the 249 genes. In total, 628 gene regulatory relationships were inferred, where the accuracy of the method was confirmed in MAP kinase cascade. The DBRF method will be a powerful tool for genome-wide gene network analysis. INTRODUCTION

Recent progress in the field of molecular biology enables us to obtain huge amounts of data. The rapidly increasing amount of known sequence data, or massive gene expression data, requires computational effort to extract information from them. So far, much attention has been focused on developing various advanced computational tools, such as for homology search, protein classification, gene clustering, and so forth.

Several significant studies have attempted to establish a method to infer a gene regulatory network from large-scale gene expression data. The gene expression data are primarily obtained as either 1) time series, or 2) steady-state data. For analyzing the time series, networks are inferred by employing various techniques (e.g., information theory (Liang et al., 1998), genetic algorithms (Morohashi and Kitano, 1999), or simulated annealing (Mjolsness et al., 1999)). One of the shortcomings of the time series approach is that it requires experimental data that are taken at very short intervals and are almost free from experimental noise. These requirements are almost impossible to meet with current techniques. On the other hand, some methods have already been proposed for inferring regulatory relationships using steady-state gene expression data. The steady-state data can be obtained by altering specific gene activities, such as knock-outing or overexpressing genes. Gene knock-outing is currently being developed on a large scale for a variety of experimental animals, such as S. cerevisiae (Winzeler et al., 1999; Hughes et al., 2000), C. elegans (Gengyo-Ando and Mitano, 2000), and Drosophila (Spradling et al., 1999), by which various gene expression profiles will be produced in a unified manner. Moreover, the discovery of RNA interference enables us to create gene knockout animals easily and is applicable to C. elegans and Drosophila (Sharp, 1999). Akutsu et al. (1998) calculated upper and lower bounds on the number of experiments that would be required if the network were Boolean. More recently, Ideker et al. (2000) proposed an inference method called predictor. The predictor method provides candidate networks represented by a Boolean network that are consistent with expression data by employing combinatorial optimization techniques. A drawback of these methods is that they assume a gene network as a Boolean network where the expression levels are represented as binary values. In general, experimental data have continuous values, and thus the data should be translated into binary data in order to apply the methods. Such translation may cause the data to lack the information needed to infer regulatory relationships. If binary data are used, even 3-state (e.g., wild-type, deletion, and overexpression) levels may be impossible to be represented, in which case the underlying inherent regulatory relationships cannot be accurately represented. In this chapter, we review the Difference-Based Regulation Finding (DBRF) method which is a gene network inference method using steadystate gene expression data (Kyoda et al., 2000). The DBRF method is applicable to expression data represented as not only binary values, but also continuous values. The chapter is organized as follows: in the next section, we review the algorithm of the DBRF method. In the third section, the performance of the DBRF method is reviewed. The performance was studied using artificial gene regulatory networks. In the fourth section, we review our application of this method to yeast DNA microarray data. In the fifth section, we discuss the advantages and characteristics of this 60

Shuichi Onami, et al.


We describe the DBRF method for inferring a gene regulatory network from the steady-state gene expression data of wild-type and deletion/overexpression mutants (Kyoda et al., 2000). An example of interaction matrix I is shown in Figure 3.1(a), which represents gene interactions. Rows of I represent the genes that regulate the genes in columns (e.g., a0 activates both a2 and a3 , and a2 represses a3 ). We assume that the data are given by an expression matrix E, a set of observed steady-state gene expression levels for all genes over all mutation experiments. An example of E is shown in Figure 3.1(b). Rows of E represent the deleted genes while columns represent the steady-state expression levels in each gene. We apply the method to the expression matrix E in order to derive the interaction matrix I . The basic procedure of the DBRF method involves two steps: 1) infer direct and indirect regulations among the genes from expression data, and 2) eliminate the indirect regulations from the above regulations to infer a parsimonious network.

a0 a0 a1 a2 a3


a2 + +

a3 + −

(a) The interaction matrix I

wt a0 − a1 − a2 − a3 −

x0 3.750 − 3.750 3.750 3.750

x1 3.750 3.750 − 3.750 3.750

x2 8.939 8.769 8.769 − 8.939

x3 0.078 0.011 0.086 5.476 −

(b) The expression matrix E

Figure 3.1 An example of the matrices which show the regulations and steady-state data of a network. (a) An interaction matrix I . (b) An expression matrix E. The values in the matrix are calculated by model equations shown in Figure 3.3(b).

Inference of a Redundant Gene Regulatory Network A simple way to determine the regulatory relationships between genes is to see the difference of expression level 1 between wild-type (wt) and mutant data. In the first step, the DBRF method derives the relationships between genes as such. The gene regulatory relationship is inferred according to the rule shown in Table 3.1. It is clear that gene a activates (represses) the expression of gene b if the expression level of gene b goes 1 ‘expression level’ is represented as absolute or relative quantities of mRNA or proteins. 61

The DBRF Method for Inferring a Gene Network from Gene Expression Data

down (up) when gene a is deleted. The computational cost of this comparison process is O(n 2 ). This process can infer not only direct gene regulations but also indirect ones. For example, the process infers a gene interaction from gene a1 to gene a3 , since the expression level x 3 is different between wt and a1− (Figure 3.1(b)). However, this interaction is an indirect gene interaction through gene a2 (Figure 3.3(a)). In the subsequent process, these indirect gene regulations are eliminated, and a parsimonious gene regulatory network is inferred. Table 3.1 Inference rule of genetic interaction between gene a and gene b from steady-state gene expression data of wild-type, single deletion and overexpression mutant.

Gene a

deletion overexpression

Expression level of gene b up down ab a→b a→b ab

Inference of a Parsimonious Gene Regulatory Network This step infers a parsimonious gene regulatory network from the redundant gene regulatory network inferred above by eliminating indirect edges (gene regulations). In order to eliminate those indirect edges, for each pair of genes, we 1) find out whether there are more than one route between those genes, 2) check whether regulatory effects (activation/inactivation) of those routes are the same, and 3) eliminate redundant routes if the effects are the same. Figure 3.2 shows the algorithm for inferring a parsimonious gene regulatory network. In order to develop 1) and 2), we modified Warshall’s algorithm (Gross and Yellen, 1999). Warshall’s algorithm is based on the transitive rule that there is an edge from ai to ak if edges from ai to a j , and from a j to ak exist. For example, if there are edges from ai to ak , from ai to a j , and from a j to ak , the algorithm finds out that there are two routes from ai to ak . Even if a route consists of more than three genes, 1) can be done using this algorithm. 2) is implemented by adding a function counting the number of negative regulations in each route to Warshall’s algorithm. The regulatory effect only depends on the parity of the number of negative regulations involved in the route (Thieffry and Thomas, 1998). For example, given two routes connecting the same pair of genes, the regulatory effects of those two routes are the same if the parities of that number are the same. The number of negative regulations is counted in each route found in 1), and groups of routes whose regulatory effects are the same, are detected in the algorithm. For 3), we define that if there 62

Shuichi Onami, et al.

procedure var

input: output: tn:

an n-node gene regulatory network G with node a1 , a2 , ..., an . the transitive closure of gene regulatory network G. total number of negative regulations.

begin initialize gene regulatory network G 0 to be network G. for i = 1 to n do for j = 1 to n do if (a j , ai ) is an edge in network G i−1 for k = 1 to n do if (ai , ak ) is an edge in network G i−1 tn = (a j , ai )negat ive num. + (ai , ak )negat ive num. if (a j , ak )negat ive num. is even, and tn is even. eliminate edge (a j , ak ) to G i−1 . (a j , ak )negat ive num. = tn. if (a j , ak )negat ive num. is odd, and tn is odd. eliminate edge (a j , ak ) to G i−1 . (a j , ak )negat ive num. = tn. return gene regulatory network G n end Figure 3.2 An algorithm for inferring a parsimonious gene regulatory network. Here let G be an n-node digraph with nodes a1 , a2 ,...,an . This algorithm constructs a sequence of digraphs, G 0 , G 1 ,...,G n , such that G 0 = G, G i is a subgraph of G i−1 , i = 1,...,n because of eliminating redundant edges subsequently. (a p , aq ) is the p-q element of the interaction matrix I . Each element of the interaction matrix I has the storage for total negative regulation number between gene p and gene q.

is more than one possible route between a given pair of genes and their regulatory effects are the same, the route consisting of the largest number of genes is the parsimonious route and the others are redundant. Thus, for each pair of genes, the number of genes in each route of the same effect is counted, and all but the one consisting of the largest number of genes are eliminated in the algorithm. The computational cost of this algorithm is O(n 3 ). COMPUTATIONAL EXPERIMENTS

Since the experimental data of deletion mutants are being produced by several yeast genome deletion projects (Winzeler et al., 1999; Hughes et al., 2000), it is reasonable to examine the performance of the DBRF method using expression data of all single gene deletions. To this end, a series of gene networks and all single gene deletion mutants for each network are simulated to generate sets of target artificial steady-state gene expression data. After generating the data sets, we apply the DBRF method to these data, and infer a gene regulatory network (Kyoda et al., 2000). 63

The DBRF Method for Inferring a Gene Network from Gene Expression Data




0.8 1


-1.3 2

(a) A network with weight values

dv 0 /dt dv 1 /dt dv 2 /dt dv 3 /dt

= 1.5g(0) − 0.2v 0 = 1.5g(0) − 0.2v 1 = 1.8g(0.8v 0 + 0.8v 1 ) − 0.2v 2 = 1.1g(2.0v 0 − 1.3v 2 ) − 0.2v 3

(b) The model equations of the network

Figure 3.3 Example of a gene regulatory network model

Network Model Here, we present the network model used for generating the artificial gene expression data. A gene regulatory network is described as a graph structure consisting of nodes an (n = 0, 1, · · · , N), directed edges between nodes with weights, and a function gn for each node. A node represents a gene, and a directed edge represents a gene regulation. The weight of a directed edge takes a positive/negative value representing activation/repression effect on the target gene. The expression level of a gene an is determined by gn , which is a nonlinear sigmoidal function reported to describe a gene expression (Mjolsness et al., 1999; Kosman et al., 1998). Thus, the expression level of gene a is described by the following equation:   dv a ab b a = Ra g W v + h −λa v a dt



where v a represents the expression level of gene a, Ra is the maximum rate of synthesis√from gene a, and g(u) is a sigmoidal function given by g(u) = (1/2)[(u/ u 2 + 1) + 1]. W ab is a connection-weight matrix element  which describes gene regulatory coefficients. b W ab v b can be replaced by  ab b b W v , allowing the equation to describe cooperative activation and repression (Mannervik et al., 1999). h a summarizes the effect of general transcription factors on gene a, and λa is a degradation (proteolysis) rate of the product of gene a. We assume that this level always takes a continuous value. Figure 3.3 shows an example of a small network with four genes. In Figure 3.3(a), each gene an is represented by a circle with gene number n. Each directed edge has an effective weight for the target gene. The model equations for each gene are shown in Figure 3.3(b). The target artificial networks were generated over a range of gene number N and maximum indegree k. For constructing a target network T with N genes and maximum indegree k, the edges were chosen randomly 64

Shuichi Onami, et al.

so that the indegree of each gene would be distributed between 1 and k. Besides, each network was generated containing cyclic-regulations, but without containing self-regulations. The parameters in the model equations and regulation type (whether each gene is regulated by gene(s) independently or cooperatively) were randomly determined. For each network, we simulated all single deletion mutants. For each of network sizes N and k, we simulated 100 target networks. Performance of the DBRF Method We analyzed the above artificial data with the DBRF method, and compared the inferred networks with the original target networks. The similarity between each inferred network and its target network was evaluated by two criteria, sensitivity and specificity. Sensitivity is defined as the percentage of edges in the target network that are also present in the inferred network, and specificity is defined as the percentage of edges in the inferred network that are also present in the target network. The results from the experiments over a range of N and k are shown in Table 3.2. As can be seen in Table 3.2, the average of specificity is always higher than that of sensitivity, and sensitivity increases in proportion to the network size N. The average of specificity is about 90% for the indegree k = 2, independent of N. The averages of sensitivity and specificity decrease in proportion to the increase of k. Comparison between Continuous-value Data and Binary Data One of the major characteristics of the DBRF method is its applicability to continuous-value expression data. To confirm the superiority of using continuous-value expression data, we applied the DBRF method to continuous-value data and binary data, and compared the results. The continuous-value expression data were translated into binary data according to a threshold; the threshold is determined as the middle value beTable 3.2 The results from the experiments over a range of N and k. Each measurement is an average over 100 simulated target networks, with standard error given in parentheses.




10 20 50 100 20 20

2 2 2 2 4 8

Total sim. edges 15.2(1.6) 30.5(0.6) 75.5(0.7) 150.4(0.6) 80.0(0.0) 117.1(11.8)

Total inferred edges 8.9(3.0) 20.6(4.9) 60.7(8.7) 133.9(10.4) 20.3(8.9) 15.1(8.6)

Num. shared edges 8.1(2.9) 18.6(4.7) 54.4(7.9) 119.2(9.5) 17.0(7.4) 9.6(6.1)



53.1% 61.1% 72.1% 79.2% 21.3% 8.1%

90.4% 90.6% 89.8% 89.1% 84.1% 61.3%

The DBRF Method for Inferring a Gene Network from Gene Expression Data

tween the minimum expression level x min (which is zero, because all expression data over single deletion mutant are given) and the maximum expression level x max . The results from the experiments over a range of N and k are shown in Table 3.3. Both sensitivity and specificity, in the case of the continuous-value data, were much higher than those in the case of the binary data. Table 3.3 The results from continuous-value and binary expression data over a range of N and k. Each measurement is an average over 100 simulated target networks. N


10 20 50 100 20 20

2 2 2 2 4 8

Continuous-value raw data sensitivity specificity 53.1% 90.4% 61.1% 90.6% 72.1% 89.8% 79.2% 89.1% 21.3% 84.1% 8.1% 61.3%

Binary translated data sensitivity specificity 20.3% 58.6% 20.9% 61.7% 22.2% 63.1% 22.9% 65.3% 9.7% 60.1% 6.4% 47.3%

Comparison with the Predictor Method The predictor method is the most recently reported gene network inference method for steady-state data (Ideker et al., 2000), and thus is considered to be the most powerful method. Therefore, we compared the performance between the DBRF method and the predictor method. The predictor method is designed to analyze binary data, and is not applicable to continuous-value data. Thus, the predictor method was applied to binary data translated from the original continuous-value data as described above, whereas the DBRF method was applied to the original data. The results show that the performance of the DBRF method is superior to that of the predictor method (Table 3.4). In the case of N = 20, k = 8, although the sensitivity of the predictor method is slightly higher than that of the DBRF method, the difference is not significant (P